U.S. patent application number 13/858041 was filed with the patent office on 2014-05-22 for real-time monocular structure from motion.
This patent application is currently assigned to NEC Laboratories America, Inc.. The applicant listed for this patent is NEC Laboratories America, Inc.. Invention is credited to Manmohan Chandraker, Shiyu Song.
Application Number | 20140139635 13/858041 |
Document ID | / |
Family ID | 50274054 |
Filed Date | 2014-05-22 |
United States Patent
Application |
20140139635 |
Kind Code |
A1 |
Chandraker; Manmohan ; et
al. |
May 22, 2014 |
REAL-TIME MONOCULAR STRUCTURE FROM MOTION
Abstract
Systems and methods are disclosed for multithreaded navigation
assistance by acquired with a single camera on-board a vehicle;
using 2D-3D correspondences for continuous pose estimation; and
combining the pose estimation with 2D-2D epipolar search to
replenish 3D points.
Inventors: |
Chandraker; Manmohan; (Santa
Clara, CA) ; Song; Shiyu; (Princeton, NJ) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
NEC Laboratories America, Inc.; |
|
|
US |
|
|
Assignee: |
NEC Laboratories America,
Inc.
Princeton
NJ
|
Family ID: |
50274054 |
Appl. No.: |
13/858041 |
Filed: |
April 6, 2013 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
61725733 |
Nov 13, 2012 |
|
|
|
61701877 |
Sep 17, 2012 |
|
|
|
Current U.S.
Class: |
348/46 |
Current CPC
Class: |
G06T 2207/30252
20130101; H04N 13/204 20180501; G06T 7/579 20170101; H04N 13/20
20180501; G06T 7/74 20170101; G06T 2207/30244 20130101 |
Class at
Publication: |
348/46 |
International
Class: |
H04N 13/02 20060101
H04N013/02 |
Claims
1. A method for vehicular navigation, comprising acquiring images
with a single camera on-board a vehicle; using 2D-3D
correspondences for continuous pose estimation on a multithreaded
processor; combining the pose estimation with 2D-2D epipolar search
to replenish 3D points with the multithreaded processor to
determine visual odometry; and applying the visual odometry for
driving the vehicle.
2. The method of claim 1, comprising pose-guided matching to
provide fast 3D-2D correspondences.
3. The method of claim 1, comprising performing epipolar
constrained search to produce per-frame 2D-2D correspondences.
4. The method of claim 1, comprising performing vanishing point
detection to hypothesize a feature match search window along one or
more radial lines from the VP for pruning mismatches due to
repeated features.
5. The method of claim 1, comprising validating each 3D point
against all frames in real-time, performing local bundle adjustment
to refine cameras and 3D points, and collecting and refining 3D
points from an epipolar thread.
6. The method of claim 1, comprising inserting new 3D points in a
main thread.
7. The method of claim 1, comprising collecting and refinding to
allow bundle adjustment using long tracks.
8. The method of claim 1, comprising performing real-time global
bundle adjustment in a thread-safe architecture with real-time pose
estimation.
9. The method of claim 1, comprising scale correcting by combining
scale estimates from 3D points and planar homography mappings.
10. The method of claim 1, comprising estimating a camera height
from the ground plane to solve scale ambiguity in monocular
SFM.
11. The method of claim 1, comprising using a 1-point RANSAC method
to estimate a height of a best-fitting plane to scattered 3D points
that are hypothesized to lie on the ground.
12. The method of claim 1, comprising using a homography-guided
dense stereo process to recover height and normal of a best fit
ground plane.
13. The method of claim 1, comprising detecting a road lane in
real-time as another cue for ground plane height and for driver and
vehicle safety applications.
14. The method of claim 1, comprising applying a Kalman filter to
combine cues for one or more ground plane height estimates.
15. The method of claim 1, comprising performing dense guided
stereo estimation, where for a hypothesized value of (h,n), a
stereo cost function determines a homography mapping between frames
k and k+1 as G = R + 1 h tn T , ##EQU00006## where pixels in frame
k+1 are mapped to frame k and a sum of absolute differences (SAD)
is computed over bilinearly interpolated image intensities.
16. The method of claim 1, wherein for each homography mapping,
determining a SAD score corresponding to the road region using
bilinearly interpolated image intensities and consider the value
s=1-.eta..sup.-SAD.
17. The method of claim 1, comprising performing triangulation on
3D points with matched SIFT descriptors between frames k and k+1,
computed within a region of interest.
18. The method of claim 1, comprising determining height estimates
with a Kalman filter and training the Kalman filter to determining
covariances.
19. The method of claim 18, comprising estimating the ground plane
with techniques j=1, . . . , m, each with observation covariance
U.sub.j and with
U.sub.k.sup.-1=.SIGMA..sub.j=1.sup.mU.sub.j,k.sup.-1, determining
fusion equations at time instant k as z k = U k j = 1 m U j , k - 1
z j , k , H k = U k j = 1 m U j , k - 1 H k . ##EQU00007##
20. A system for navigation, comprising a single camera on-board a
vehicle; and a multi-threaded processor coupled to the camera, the
multi-threaded processor using 2D-3D correspondences for continuous
pose estimation and combining the pose estimation with 2D-2D
epipolar search to replenish 3D points and applying visual odometry
for driving the vehicle.
21. The system of claim 20, comprising code for scale correcting by
combining scale estimates from 3D points and planar homography
mappings.
22. The system of claim 20, comprising code for pose-guided
matching to provide fast 3D-2D correspondences.
23. The system of claim 20, comprising code to perform epipolar
constrained search to produce per-frame 2D-2D correspondences.
24. The system of claim 20, comprising code for estimating a camera
height from the ground plane to solve scale ambiguity in monocular
SFM.
25. The system of claim 20, comprising code for detecting a road
lane in real-time as another cue for ground plane height and for
driver and vehicle safety applications.
26. The system of claim 20, comprising code for applying a Kalman
filter to combine cues for one or more ground plane height
estimates.
Description
[0001] This application claims priority to Provisional Application
Ser. Nos. 61/701,877 filed on Sep. 17, 2012 and 61/725,733 filed on
Nov. 13, 2012, the content of which is incorporated by
reference.
BACKGROUND
[0002] The present invention relates to monocular structure from
motion (SFM).
[0003] Autonomous driving faces unique challenges as a difficult
corner case for SFM. Traditional systems for unordered image
collections are inapplicable in such situations, since
fundamentally, forward motion is an ill-posed situation. This is
compounded by the fact that high vehicle speeds lead to rapidly
changing imagery, so successful indoor systems like that rely on
repeated observations of the same scene elements also break down.
Further, the timing demands on autonomous driving systems are
extremely stringent--a reliable camera pose is expected at every
frame, with no luxury of delayed verifications or bundle
adjustments.
[0004] While stereo Simultaneous Localization And Mapping (SLAM)
systems routinely achieve high accuracy and real-time performance,
the challenge remains daunting for monocular ones. Yet, monocular
systems are attractive for the automobile industry since they are
cheaper and the calibration effort is lower. Costs of consumer
cameras have steadily declined in recent years, but cameras for
practical visual odometry in automobiles are expensive since they
are produced in lesser volume, must support high frame rates and be
robust to extreme temperatures, weather and jitters.
[0005] The challenges of monocular visual odometry for autonomous
driving are both fundamental and practical. For instance, it has
been observed empirically and theoretically that forward motion
with epipoles within the image is a "high error" situation for SFM.
Vehicle speeds in outdoor environments can be high, so even with
cameras that capture imagery at high frame rates, large motions may
occur between consecutive frames. This places severe demands on an
autonomous driving visual odometry system, necessitating extensive
validation and refinement mechanisms that conventional systems do
not require. The timing requirements for visual odometry in
autonomous driving are equally stringent--a pose must be output at
every frame in a fixed amount of time. For instance, traditional
systems may produce a spike in timings when keyframes are added, or
loop closure is performed.
SUMMARY
[0006] In one aspect, systems and methods are disclosed for
multithreaded visual odometry by acquired with a single camera
on-board a vehicle; using 2D-3D correspondences for continuous pose
estimation; and combining the pose estimation with 2D-2D epipolar
search to replenish 3D points.
[0007] In another aspect, an accurate, robust and real-time
large-scale SFM system for real-world autonomous outdoor driving
application is disclosed. The system uses multithreaded
architectures for SFM that allow handling large motions and rapidly
changing imagery for fast-moving vehicles. The system includes
parallel epipolar search for extensive validation of feature
matches over long tracks and a keyframe architecture that allows
insertion at low cost. This allows robust operation of the system
at 30 fps on the average, with output guaranteed at every frame
within 50 ms. To resolve the scale ambiguity of monocular SFM, the
system estimates the height of the ground plane at every frame.
Cues for ground plane estimation include triangulated 3D points and
plane-guided dense stereo matching. These cues are combined in a
flexible Kalman filtering framework, which is trained rigorously to
operate with the correct empirical covariances.
[0008] Advantages of the above aspect may include one or more of
the following. The system makes judicious use of a multithreaded
design to ensure that motion estimates (and the underlying
structure variables) become available only after extensive
validation with long-range constraints and thorough bundle
adjustments, but without delay. Thus, the system is optimized for
worst-case timing scenarios, rather than the average-case
optimization for most traditional systems. In particular, the
multithreaded system produces pose outputs in at most 50 ms,
regardless of whether a keyframe is added or scale correction
performed. The average frame rate of the system is much higher, at
above 30 fps.
[0009] The system provides a real-time, accurate, large-scale
monocular visual odometry system for real-world autonomous outdoor
driving applications. The architecture of the system addresses the
challenge of robust multithreading even for scenes with large
motions and rapidly changing imagery. The design is extensible for
three or more parallel CPU threads. The system uses 3D-2D
correspondences for robust pose estimation across all threads,
followed by local bundle adjustment in the primary thread. In
contrast to prior work, epipolar search operates in parallel in
other threads to generate new 3D points at every frame. This
significantly boosts robustness and accuracy, since only
extensively validated 3D points with long tracks are inserted at
keyframes. Fast-moving vehicles also necessitate immediate global
bundle adjustment, which is triggered by the instant keyframe
design in parallel with pose estimation in a thread-safe
architecture. To handle inevitable tracking failures, a recovery
method is provided. Scale drift is corrected using a mechanism that
detects (rather than assumes) local planarity of the road by
combining information from triangulated 3D points and the
inter-image planar homography. The system is optimized to output
pose within 50 ms in the worst case, while average case operation
is over 30 fps. Evaluations are presented on the challenging KITTI
dataset for autonomous driving, where the system achieves better
rotation and translation accuracy than other systems.
BRIEF DESCRIPTION OF THE DRAWINGS
[0010] FIG. 1 shows an exemplary multithreaded architecture with
2D-3D correspondences for continuous pose estimation.
[0011] FIG. 2 shows an exemplary pose-guided matching module to
provide fast 3D-2D correspondences.
[0012] FIG. 3 shows an exemplary epipolar constrained search.
[0013] FIG. 4 shows an exemplary local bundle module.
[0014] FIG. 5 shows an exemplary multithreaded keyframe
architecture to handle insertion of new 3D points in the main
thread.
[0015] FIG. 6 shows an exemplary collection and refining module
that allows bundle adjustment using long tracks.
[0016] FIG. 7 shows an exemplary real-time global bundle adjustment
module in a thread-safe architecture with real-time pose
estimation.
[0017] FIG. 8 shows in more details a process for ground plane
estimation with 3D points 301.
[0018] FIG. 9 shows an exemplary process for ground plane
estimation with guided dense stereo.
[0019] FIG. 10 shows an exemplary process for real-time lane
detection.
[0020] FIG. 11 shows an exemplary process for scale correction.
DESCRIPTION
[0021] While stereo-based SFM is commonly available in commercial
products, the lack of a fixed baseline renders monocular SFM a far
more daunting challenge. Yet, the monocular SFM is cost-effective
in mass production, thus, an attractive proposition for the
autonomous driving industry.
[0022] FIG. 1 shows a monocular SFM embodiment with performance
equal to or better than that of stereo SFM systems. To meet the
competing demands of accuracy and timing, the system design entails
a multithreaded architecture. Camera poses are determined using
2D-3D correspondences, where a dedicated epipolar search thread
continually updates candidate 3D points. A keyframe design ensures
that the epipolar search supplies long range constraints to a
bundle adjustment module operating in parallel. This allows the
system to utilize only points that have undergone extensive
validation and output refined poses in real-time, at every
frame--in contrast to browsing systems like PTAM that perform
epipolar search on-demand and delayed refinement.
[0023] The monocular SFM handles scale drift, which does not occur
in fixed baseline stereo SFM. The most popular SLAM approach for
scale correction is loop closure. However, it is impractical for
autonomous driving applications, since one cannot rely on the
presence of loops in general road conditions. Moreover,
applications like scene understanding and driver safety cannot
compromise on accuracy or timing between loops closures. The system
achieves a high degree of robustness through an extensive,
principled scale correction mechanism that relies on real-time
estimates of camera height from the ground plane.
[0024] Ground plane estimation is itself a challenging problem, due
to the lack of reliable texture on the road. The system counters
this by combining cues from multiple methods in a rigorously
operated, but flexible, Kalman filter framework. The cues used are
triangulated 3D points and a dense stereo matching in real-time to
find the optimal planar homography mapping between the road in two
frames. While each of these methods is unreliable on its own, a
well-designed Kalman filter successfully combines them to produce
robust estimates. This allows the system to greatly improve and
complete the KITTI test sequences with low errors.
[0025] An innovation central to the excellent performance of the
data fusion is the correct computation of observation
covariances--an aspect often overlooked by other SFM systems that
use Kalman filters. This is achieved through an extensive training
procedure over long distances of real-world driving sequences, that
learns relationships between variance in ground plane estimation to
the underlying variables for each method. Thus, during the actual
operation of the filter, the system predicts meaningful empirical
covariances that are combined to produce a reliable ground plane
estimate.
[0026] FIG. 1 shows an exemplary multithreaded architecture with
2D-3D correspondences for continuous pose estimation. One
embodiment is a multithreaded structure-from-motion (SFM)
architecture 100. The steady-state architecture 100 includes a pose
local bundle adjustment (LBA) system 101/103 and an epipolar search
unit 102. The multithreaded architecture uses 2D-3D correspondences
for continuous pose estimation in 101/103. This is combined with
2D-2D epipolar search to replenish 3D points. This architecture
allows camera pose estimation and 3D reconstruction using fast
moving imagery acquired with a single camera on-board a vehicle,
thereby enabling autonomous driving applications.
[0027] The output of the pose LBA is provided to a keyframe and
keyframe+1 unit 201/202. Unit 201 provides a collection and
refinding mechanism allows bundle adjustment using long tracks, as
opposed to two-view estimations for previous works, while unit 202
handles the keyframe+1 where real-time global bundle adjustment is
conducted in a thread-safe architecture with real-time pose
estimation.
[0028] The system also includes a ground plane estimation unit 300
using 3D points 301 and guided dense stereo information 302. The
system also performs lane detection 400 with real time line
matching. The result from ground plane estimation and lane
detection are provided to a scale recovery 500 that uses a Kalman
filter for cue combination in one embodiment.
[0029] The multithreaded architecture allows elegant extension to
as many threads as desired. Besides speed advantages,
multithreading also greatly contributes to the accuracy and
robustness of the system. As an example, consider the epipolar
contrained search. A single-thread version of a system that relies
on 2D-3D correspondences might update its stable point set by
performing an epipolar search in the frame preceding a keyframe.
However, the support for the 3D points introduced by this mechanism
is limited to just the triplet used for the circular matching and
triangulation. By moving the epipolar search to a separate thread
and performing the circular matching at every frame, the system may
supply 3D points with tracks of length up to the distance from the
preceding keyframe. Clearly, the set of long tracks provided by the
epipolar thread in the multithread system is far more likely to be
free of outliers.
[0030] The above multithreaded architecture for monocular SFM is
designed to meet the challenges of autonomous driving. Real-time
operation is achieved at 30 fps, with guaranteed refined pose
output within 50 ms. Robust scale correction is done by ground
plane estimation using multiple cues in a flexible Kalman filter
framework. Data fusion is achieved with rigorous, data-driven
computation of observation covariances for each cue. The system
achieves accuracy that outperforms or rivals state-of-the-art
stereo systems in rotation and translation.
[0031] FIG. 2 shows one exemplary pose-guided matching module 100
(FIG. 1) to provide fast 3D-2D correspondences. At steady state,
the system has access to a stable set of 3D points. The poses of
the previous frames have undergone multiple non-linear
optimizations and are considered accurate. Each frame is processed
by three modules: the pose module estimates the pose of the current
frame and occupies all available threads; 2) the epipolar update
module uses epipolar constrained search (ECS) and triangulation T
to prepare new 3d points to be added; and 3) the local bundle
adjustment (LBA) unit performs non-linear optimization over a
sliding window of a few previous frames and takes place in the main
thread.
[0032] Pose-guided matching with fast 3D-2D correspondences is
supported by the architecture of FIG. 2 for every steady state
frame. The modules are depicted in their multithreading
arrangement, in correct synchronization order but not to scale.
Camera images are provided to a detection module, then provided to
a PGM unit for pose-guided matching, and then to a perspective
n-point (PnP) pose estimation unit. The outputs of the PnP unit are
provided to the LBA unit, then to a finding unit R and an update
motion model U. The output of the PnP unit are also provided to an
ECS unit for epipolar constrained search and then to a
triangulation unit T.
[0033] To initialize, the system extracts FAST corners with ORB
descriptors and matches between consecutive frames using locality
sensitive hashing (LSH). With sufficient baseline (around 5
frames), a set of 3D points is initialized by relative pose
estimation, triangulation and bundle adjustment. Each frame during
initialization is processed within 10 ms.
[0034] At steady state, the system has access to a stable set of at
least N.sub.s 3D points that have undergone extensive bundle
adjustment in prior frames (for N.sub.s=100). The preceding poses
have also undergone multiple non-linear refinements, so can be
considered highly accurate. The system architecture at every frame
in steady state operation is illustrated in FIG. 1.
[0035] Around 2000 FAST corners with Shi-Tomasi filtering are
extracted from a typical outdoors image and ORB descriptors are
computed. Using the pose of the previous frame, the pose of the
current frame is predicted, assuming constant velocity. The system
explicitly computes the camera pose at each frame using
correspondences, the motion model is only used as guidance to
expedite matching. The existing set of stable 3D points are
projected into the image using the predicted pose and the ORB
descriptor for each is compared to those within a square of side
2r.sub.s pixels (for example r.sub.s=15). Given these 2D-3D point
correspondences, the system computes the actual camera pose using
perspective n-point (PnP) pose estimation in a RANSAC framework.
The particular implementation used is EPnP with a model size of 4
points. The RANSAC pose with the largest consensus set is refined
using a Levenberg-Marquardt nonlinear optimization.
[0036] The system can easily handle other choices for matching, in
particular, it has achieved similar results using normalized
cross-correlation (NCC) instead of ORB. But associating a
descriptor like ORB with a 3D point can have ancillary
benefits.
[0037] If the set of 3D points in application scenes remains
unchanged, the pose module is enough to maintain camera pose for
extended periods. However, unlike small workspace environments,
scene points rapidly move out of view in outdoor applications and
candidate sets of points usable for pose estimation must be
continually updated in a thread of their own (rather than on-demand
like PTAM).
[0038] For every feature f.sub.0 in the most recent keyframe at
location (x.sub.0, y.sub.0), a square of side 2r.sub.e centered at
(x.sub.0+.DELTA.x, y.sub.0+.DELTA.y) in frame n is considered. The
displacement (.DELTA.x, .DELTA.y) is computed based on the distance
of (x.sub.0, y.sub.0) from the vanishing point, which is a strong
cue in highway sequences. This vastly improves the feature matching
performance when the vehicle is moving at high speeds.
[0039] The ORB descriptors for all FAST corners within the
intersection of this square with a rectilinear band p pixels wide
centered around the epipolar line corresponding to f.sub.0 in frame
n are compared with the descriptor for f.sub.0. The closest Hamming
match, f.sub.n, is found. A match is accepted only if there is also
a matching feature in frame f.sub.n-1. Only two matches, (0, n) and
(n-1, n), are needed at frame n, since matches (0, n-1) have
already been computed for f.sub.n-1. Parameter values used are p=3
and r.sub.e=min{1200P.omega.P.sup.2, 10}, where .omega. is the
differential rotation between frames.
[0040] The features matched in frame n are triangulated with
respect to the most recent keyframe, which takes about 2 ms per
frame. The 3D point is back-projected into all frames 1, . . . ,
n-1 and retained only if a matching ORB descriptor is found within
a tight square of side 2r.sub.b pixels (r.sub.b=3). A two-view
triangulation suffices instead of a more expensive multiview
alternative, since the long tracks built by a dedicated epipolar
search module crucially ensure that only the most reliable 3D
points are inserted at keyframes, as described next.
[0041] A sliding window bundle adjustment operates in a parallel
thread with the epipolar update module. A frame cache is used for
storing feature locations, descriptors, and camera poses for the
most recent N frames (N=10). Another new feature of the system is
that it forces the previous keyframe to be in the local bundle
cache if it is not already present. Since a criterion for keyframes
to be added is that the pose has changed significantly, adding the
previous keyframe allows the system to produce stable pose results
even when the vehicle is not moving (or moving slowly). In the LBA
module, the system is also given a chance to re-find 3D points
temporarily lost due to artifacts like blur or specularities. The
publicly available SBA package [13] is used for bundle adjustment.
Timings for epipolar update and local bundle adjustment are
summarized in Table 1.
TABLE-US-00001 TABLE 1 Pose, epipolar update and local bundle
module timings in steady state, with the latter two designed to
operate in parallel. Module Operation Timing 4*Pose FAST +
Shi-Tomasi 1 ms ORB descriptor 5 ms Pose-guided matching 1 ms PnP
(RANSAC, 500 15 ms iter.) Nonlin. pose refine 1 ms 2*Epi. Update
Constrained search 10-15 ms Triangulation 1-3 ms 3*Local Bundle
Windowed bundle adj. 10-20 ms Re-find 3D points 1 ms Update motion
model 0 ms
[0042] The pose module 101 is shown in FIG. 3. As shown in FIG. 3,
the pose module has the following functions: detection of features
and ORB descriptors; pose-guided matching (PGM); and pose
estimation (PnP).
[0043] If the application scenario involves scenes where the same
set of 3D points is viewed for extended periods of time, then the
pose module by itself would be sufficient to maintain the camera
pose. However, in outdoor applications like autonomous navigation,
3D scene points rapidly move out of view within a few frames. Thus,
the stable set of points used for pose computation must be
continually updated, which is the task entrusted to the epipolar
search module 102 of FIG. 1.
[0044] FIG. 4 shows in more details the epipolar constrained search
module 102. As depicted in FIG. 4, the epipolar search module is
parallelized across two threads and follows pose estimation at each
frame. The mechanism for epipolar search is illustrated in FIG. 2.
Let the most recent prior keyframe be frame 0. After pose
computation at frame n, for every feature f.sub.0 in the keyframe
at location (x.sub.0, y.sub.0), the system considers a square of
side 2r.sub.e centered at (x.sub.0, y.sub.0) in frame n. The system
considers the intersection region of this square with a rectilinear
band p pixels wide, centered around the epipolar line corresponding
to f.sub.0 in frame n. The ORB descriptors for all FAST corners
that lie within this intersection region are compared to the
descriptor for f.sub.0. The closest match, f.sub.n, is found in
terms of Hamming distance. This epipolar matching procedure is also
repeated by computing the closest match to f.sub.n in frame n-1,
call it f.sub.n-1. A match is accepted only if f.sub.n-1 also
matches f.sub.0. Note that only two sets of matches with respect to
frames (0, n) and (n-1, n) must be computed at the frame n, since
the matches between (0, n-1) have already been computed at frame
n-1.
[0045] In one embodiment, the parameter r.sub.e is automatically
determined by the size of the motion, the system uses
r.sub.e=min{1200.parallel..omega..parallel..sup.2, 10}, where
.omega. is the differential rotation between frames n-1 and n.
Since pose estimates are highly accurate due to continuous
refinement by bundle adjustment, epipolar lines are deemed accurate
and a stringent value of p=3 can be used to impose the epipolar
constraint. The Hamming distance computation for 256-bit ORB
descriptors in a region of interest is performed as a block, with a
fast SSE implementation. To rapidly search for features that lie
within the above region of interest, the detected features in an
image are stored in a lookup table data structure. The key into the
table is the column index of the feature and within each bucket,
features are stored in sorted row order. Across two threads, this
allows circular matching for a triplet of images, with up to 500
features in each, in 10-15 ms. As opposed to brute-force searching,
the lookup table results in speedups by up to a factor of 10,
especially in the autonomous driving application where the images
traditionally have wide aspect ratios (to cover greater field of
view while limiting uninformative regions like sky).
[0046] The features that are circularly matched in frame n are
triangulated with respect to the most recent keyframe (frame 0).
This two-view triangulation requires approximately 2 ms per frame.
The reconstructed 3D point is back-projected in all the frames 1, .
. . , n-1 and is retained only if a matching ORB descriptor is
found within a very tight square of side 2r.sub.b pixels (set
r.sub.b=3). This acts as a replacement for a more accurate, but
expensive, multiview triangulation and is satisfactory since
epipolar search produces a large number of 3D points, but only the
most reliable ones may be used for pose estimation. However, these
3D points are not added to the stable point cloud yet. For that
they must first undergo a local bundle adjustment and be collected
by the main thread at a keyframe, which are aspects explained in
the following sections.
[0047] The epipolar constrained search is implemented on a thread
of its own to produce per-frame 2D-2D correspondences. For current
frame n, only 3D points that are validated against all frames 1 to
n-1 are retained. Only persistent 3D points that survive for
greater than L frames may be collected by the next keyframe.
[0048] The advantage of the above approach is that the system can
construct long tracks, so when new 3D points are inserted, they are
guaranteed to be accurate. To boost robustness, each 3D point is
validated against all the frames in real-time, while prior systems
could only do this in computational off-cycles.
[0049] If the application scenario involves scenes where the set of
3D points being viewed remains unchanged, then the pose module by
itself would be sufficient to maintain the camera pose over
extended periods. However, in outdoor applications like autonomous
navigation, 3D scene points rapidly move out of view within a few
frames. Thus, the stable set of points used for pose computation
must be continually updated, which is the task entrusted to the
epipolar search module.
[0050] FIG. 4 shows an exemplary local bundle module 103. The local
bundle adjustment refines cameras and 3D points. Data structures
are implemented to collect and refine 3D points from the epipolar
thread.
[0051] To refine camera poses and 3D points incorporating
information from multiple frames, the system implements a sliding
window local bundle adjustment. The key data structure is the local
bundle cache, which is composed of a frame cache and a match cache.
The frame cache stores feature locations, descriptors and camera
poses from the most recent N frames. It also stores images for
those N frames, for display and debugging purposes. In the system,
N=10. The match cache is a list of tables, one element
corresponding to each frame. The key into the table is the identity
of a 3D point visible in the frame and the stored entries are the
identities of the corresponding 2D features in various frames.
[0052] The local bundle adjustment module also has other functions.
After bundle adjustment, the system has a chance to re-find lost 3D
points using the optimized pose. Since the system spends
considerable effort in maintaining a high-quality set of 3D points
for pose computation, it is worthwhile to incur a small overhead to
recover any temporarily lost ones (due to image artifacts like
blur, specularities or shadows). In fact, a stable 3D point is
permanently discarded only when its projection using the current
pose falls outside the image boundaries. Since the bundle adjusted
pose is highly accurate, the system can perform re-finding by
matching ORB descriptors on FAST corners within a very tight square
of side 2r.sub.f pixels (with r.sub.f=10). This ensures re-finding
is rapidly achieved within 1 ms.
[0053] One implementation uses the publicly available SBA package
for bundle adjustment. In parallel, the motion model for predicting
the pose of the next frame is also updated in this module. The
timings for the parallel epipolar update and local bundle
adjustment modules are summarized in Table 2.
TABLE-US-00002 TABLE 2 Epipolar update and local bundle timings in
steady state (parallel modules) Module Operation Timing 2*Epipolar
Constrained search 10-15 ms Update Triangulation 1-3 ms 3*Local
Bundle Windowed bundle 10-20 ms adjustment Re-find 3D points 1 ms
Update motion model 0 ms
[0054] FIG. 6 shows an exemplary multithreaded keyframe
architecture 201 to handle insertion of new 3D points in the main
thread. The system architecture for keyframes is similar to that of
FIG. 1, with the addition of a collection and refinding module C+R.
It collates persistent 3D points tracked over at least frames in
the epipolar thread and re-finds them in the current frame using
the output of the pose module. The LBA is now different from that
for steady state, since its cache has been updated with 3D points
and their corresponding 2D locations in all the relevant frames on
the epipolar thread.
[0055] The system cannot maintain steady state indefinitely, since
3D points are gradually lost due to tracking failures or when they
move out of the field of view. The latter is an important
consideration in "forward moving" systems for autonomous driving
(as opposed to "browsing" systems such as PTAM), so the role of
keyframes is very important in keeping the system alive. The
purpose of a keyframe is threefold: [0056] Collect 3D points with
long tracks from the epipolar thread, refine them with local bundle
adjustment and add to the set of stable points in the main thread.
[0057] Trigger global bundle adjustment based on previous few
keyframes that refines 3D points and keyframe poses. [0058] Provide
the frame where newly added 3D points "reside".
[0059] The modules that define operations at a keyframe are
illustrated in FIG. 6. The pose module remains unchanged from the
steady state. It is followed by a collection stage, where 3D points
triangulated at each frame in the epipolar thread are gathered by
the main thread. Only persistent 3D points that stem from features
matched over at least L frames are collected (our circular matching
for epipolar search ensures this is easily achieved by seeking 3D
points only from at least L frames after the previous keyframe).
Note that this mechanism imposes two necessary conditions for a
point to be considered for inclusion into the stable set--it must
be visible in at least two keyframes and must be tracked over at
least L frames. While stringent, these conditions inherently
enhance the chances that only reliable 3D points are added into the
main thread. In the system, L=3 regardless of environment.
[0060] The collected 3D points must reside on a keyframe for all
subsequent operations, so a re-finding operation is performed by
projecting them using the estimated pose for the frame and finding
the best ORB match in a circular region of radius 10 pixels. Now
the existing stable 3D points, the collected 3D points from the
epipolar thread, their projections in all the frames within the
local bundle cache and the corresponding cameras undergo local
bundle adjustment. The bundle adjustment at keyframes differs from
steady state operation, but adding long tracks into the bundle
adjustment at keyframes is a reason the system can avoid more
expensive multiview triangulation at each frame in the epipolar
thread. The refined 3D points are now ready to be added to the
stable set.
[0061] The modules that define operations at the frame immediately
after a keyframe are illustrated in FIG. 7. The pose module
re-finds the (new) set of stable 3D points. The pose module is now
split across only two threads, in order to accommodate a global
bundle adjustment in the main thread. This bundle adjustment
involves the previous K keyframes and their associated 3D points,
in order to introduce long-range constraints to better optimize the
newly added set of 3D points. For the system, choosing K=5 allows
the global bundle adjustment to finish within 15 ms. There are two
reasons a more expensive bundle adjustment involving a much larger
set of previous keyframes (or even the whole map) is not necessary
to refine 3D points with long-range constraints. First, the imagery
in autonomous driving applications is fast moving and does not
involve repetitions, so introducing more keyframes into the global
bundle adjustment yields at best marginal benefits. Second, the
goal is instantaneous pose output rather than map-building, so even
keyframes are not afforded the luxury of delayed output. This is in
contrast to parallel systems like where keyframes may produce a
noticeable spike in the per-frame timings.
[0062] In FIG. 7, the previous K frames are provided to a GBA or
global bundle adjustment unit. The GBA unit usually finishes within
the time consumed by the pose module. The cache update module
reconciles the 3D points modified by both PnP and GBA, before it is
used by LBA. Following global bundle adjustment, the 3D coordinates
of all the points are updated. Note that overlapping sets of 3D
points are used by both global bundle adjustment and pose modules
in parallel, however, both may also cause this set to change (PnP
may reject 3D points that are outliers, while bundle adjustment may
move the position of 3D points). To ensure thread safety, an update
module is included that reconciles changes in the 3D point cloud
from both the prior parallel modules. The local bundle adjustment
module, which simply reads in 3D point identities, receives this
updated set for optimization based on the N frames in the local
bundle cache. In parallel with local bundle adjustment, the
epipolar search also makes use of the updated keyframe pose. While
the keyframe pose has seen a global bundle adjustment, the pose of
the subsequent frame has not. This does not cause any inconsistency
in practice since poses tend to be much more stable than points--a
camera is constrained by hundreds of points, but a point is visible
only in a few cameras. Thereafter, the system resumes steady-state
operation until the next keyframe, unless a recovery or firewall
condition is triggered. The following sections explain those
concepts in detail.
[0063] FIG. 8 shows in more details a process for ground plane
estimation with 3D points 301. FIG. 9 shows an exemplary process
for ground plane estimation with guided dense stereo. FIG. 10 shows
an exemplary process for real-time lane detection, and FIG. 11
shows an exemplary process for scale correction.
[0064] Since scale information is lost in monocular SFM, an
integral component of monocular visual odometry is scale
correction. Traditional methods for scale correction include loop
closure and estimating the height of the ground plane. For
autonomous driving applications, since loop closure is an unlikely
scenario, the latter method is used. The KITTI dataset provides the
ground truth mounting of the camera height as 1.70 meters, with a
camera pitch angle of .theta.=-0.03 radians. Multiple methods are
used for ground plane estimation and a principled approach is used
to combine the cues using a Kalman filter, whose process
covariances are rigorously learned from training data.
[0065] A Plane-guided Dense Stereo technique is used. The system
assumes a region in the foreground (middle fifth of the lower third
of the image) to be the road plane. For a hypothesized value of
(h,n), the stereo cost function computation determines the
homography mapping between frames k and k+1 as
G = R + 1 h tn T . ##EQU00001##
Pixels in frame k+1 are mapped to frame k and the sum of absolute
differences (SAD) is computed over bilinearly interpolated image
intensities. A Nelder-Mead simplex routine is used to estimate the
(h, n) that minimize this cost function. Note that the optimization
only involves the three variables h, n.sub.1 and n.sub.3, since
.parallel.n.parallel.=1.
[0066] Triangulated 3D Points are also used. The system considers
matched SIFT descriptors between frames k and k+1, computed within
the above region of interest (ORB descriptors are not powerful
enough for the low texture in this region). To fit a plane through
the triangulated 3D points, one option is to estimate (h, n) using
a 3-point RANSAC for plane-fitting, however, better results are
obtained by assuming the camera pitch to be fixed at .theta.. For
every triangulated 3D point, the height difference .DELTA.h is
computed with respect to every other point. The estimated ground
plane height is the height of the point that maximizes the
score
q = i = 1 n - .mu..DELTA. h , where .mu. = 50. ( 1 )
##EQU00002##
[0067] To combine the height estimates from various methods, a
natural framework is a Kalman filter. The system performs a
rigorous training to compute the involved covariances. The Kalman
filter model of state evolution is given by
x.sub.k=Ax.sub.k-1+w.sub.k-1, p(w):N(0,Q), (2)
z.sub.k=Hx.sub.k+v.sub.k-1, p(v):N(0,U), (3)
where x is the state variable, z the observation, while Q and U are
the covariances of the process and observation noise, respectively,
that are assumed to be zero mean multivariate normal distributions.
In one case, state variable is simply the equation of the ground
plane, thus, x=(n.sub.1, n.sub.2, n.sub.3, h).sup.T. Since
.parallel.n.parallel.=1, n.sub.2 is determined by n.sub.1 and
n.sub.3 and the observation is z=(n.sub.1, n.sub.3, h).sup.T. Thus,
the state transition matrix and the observation model are given
by
A = [ R t 0 T 1 ] T , H = [ 1 0 0 0 0 0 1 0 0 0 0 1 ] . ( 4 )
##EQU00003##
[0068] If methods j=1, . . . , m are used for estimating the ground
plane, each with its observation covariance U.sub.j Then, with
U.sub.k.sup.-1=.SIGMA..sub.j=1.sup.mU.sub.j,k.sup.-1, the fusion
equations at time instant k are
z k = U k j = 1 m U j , k - 1 z j , k , H k = U k j = 1 m U j , k -
1 H k . ( 5 ) ##EQU00004##
Meaningful estimation of U.sub.k at every frame, with the correctly
proportional U.sub.j,k for each cue, is essential to the success of
a Kalman filter-based cue combination. In the following, a
comprehensive training procedure estimates the various observation
covariances.
[0069] For Dense Stereo, the system makes the approximation that
state variables are uncorrelated. The system first fixes the values
of n.sub.1 and n.sub.3 to the optimal values from dense stereo and
for a range of h, then constructs the homography mapping from frame
k to k+1, given by
R + 1 h tn T . ##EQU00005##
For each homography mapping, the system computes the SAD score
corresponding to the road region using bilinearly interpolated
image intensities and consider the value s=1-.eta..sup.-SAD, where
.eta.=1.5. A univariate Gaussian is now fit to the distribution of
s and the variance .sigma..sub.k.sup.s is recorded (a different
.sigma..sup.s is computed for h, n.sub.1 and n.sub.3).
[0070] For each frame, let e.sub.k.sup.s be the error of the ground
plane height estimated from dense stereo alone. Then, the system
considers the histogram of e.sub.k.sup.s with B=1000 bins over the
variances .sigma..sub.k.sup.s, with the bin centers positioned to
match the density of .sigma..sub.k.sup.s (that is, distribute
roughly F/B error observations within each bin). The variances
.sigma..sub.e.sup.s corresponding to the e.sup.s are computed
within each bin b=1, . . . , B and a curve is fitted to the
distribution of .sigma..sub.e.sup.s versus .sigma..sup.s.
Empirically, a straight line suffices to produce a good fit. A
similar process is repeated for n.sub.1 and n.sub.3.
[0071] During testing when the Kalman filter is in operation, the
system fits a 1D Gaussian to the homography-mapped SAD scores to
get the value of .sigma..sup.s, corresponding to h, n.sub.1 and
n.sub.3. Using the line-fit parameters estimated above, the system
can predict the value of .sigma..sub.e.sup.s. The covariance matrix
for the dense stereo method is now available as
U.sub.1=diag(.sigma..sub.e.sup.s(n.sub.1),.sigma..sub.e.sup.s(n.sub.3),.s-
igma..sub.e.sup.s(h)).
[0072] The covariance estimation for the method that uses
triangulated 3D points differs from the stereo method, since the
normal n is assumed known from the camera pitch and only the height
h is an estimated entity. During training, for various trial values
of h at frame k, the system computes the height error e.sub.k.sup.p
with respect to the ground truth and the sum q defined in (1). As
in the case of dense stereo, a histogram is computed for the with
B=1000 bins and approximately F/B observations of e.sub.k.sup.p are
recorded at each bin, centered at q.sub.b, for b=1, . . . , B.
[0073] Since n.sub.1 and n.sub.3 are fixed to values from camera
pitch angle, fixed variance estimates .sigma..sup.p(n.sub.1) and
.sigma..sup.p(n.sub.3) are computed for them, as the variance of
the errors in n.sub.1 and n.sub.3 with respect to ground truth.
During testing, the value of q is computed by the ground plane
estimation using (1) and the corresponding value of
.sigma..sup.p(h) is estimated from the above line-fit. The
covariance matrix for this method in the Kalman filter data fusion
is now available as
U.sub.1=diag(.sigma..sup.p(n.sub.1),.sigma..sup.p(n.sub.3),.sigma..sup.p(-
h)).
[0074] The multithreaded system enables large-scale, real-time,
monocular visual odometry, targeted towards autonomous driving
applications with fast-changing imagery. The multithreaded design
can boost both the speed and accuracy for handling challenging road
conditions. The system is optimized to provide pose output in
real-time at every frame, without delays for keyframe insertion or
global bundle adjustment. This is achieved through a per-frame
epipolar search mechanism that generates redundantly validated 3D
points persistent across long tracks and an efficient keyframe
architecture to perform online thread-safe global bundle adjustment
in parallel with pose computation. Further, the system is accurate
enough to require only occasional scale correction, for which an
automatic mechanism is provided that detects planarity of the
ground to compute reliable scale factors. Multithreaded bundle
adjustment can be optimized for small-sized problems that arise in
autonomous driving applications. Real-time detection of pedestrians
and cars can also be done for better handling of crowded
scenes.
* * * * *