HIGH ACCURACY MONOCULAR MOVING OBJECT LOCALIZATION

Abstract

Methods and systems for moving object localization include estimating a ground plane in a video frame based on a detected object within the video frame and monocular structure-from-motion (SFM) information; computing object pose for objects in the frame based on the SFM information using dense feature tracking; and determining a three-dimensional location for the detected object based on the estimated ground plane and the computed object pose.

Description

BACKGROUND OF THE INVENTION

Three-dimensional localization of moving objects in a video stream using only a single camera can be difficult. In particular, existing methods use sparse feature points, but this technique is difficult to use on objects such as, e.g., cars, because it is hard to establish stable feature tracks. Other existing methods triangulate object bounding boxes against a fixed ground plane, which leads to high localization errors. These existing techniques may also involve expensive inference mechanisms.

BRIEF SUMMARY OF THE INVENTION

A method for moving object localization includes estimating a ground plane in a video frame based on a detected object within the video frame and monocular structure-from-motion (SFM) information. Object pose for objects in the frame is computed based on the SFM information using dense feature tracking. A three-dimensional location for the detected object is determined based on the estimated ground plane and the computed object pose.

A system for moving object localization includes a processor configured to estimate a ground plane in a video frame based on a detected object within the video frame and monocular structure-from-motion (SFM) information, to compute object pose for objects in the frame based on the SFM information using dense feature tracking, and to determine a three-dimensional location for the detected object based on the estimated ground plane and the computed object pose.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block/flow diagram of object localization in accordance with the present principles.

FIG. 2 is a block diagram of dense three-dimensional estimation in accordance with the present principles.

FIG. 3 is a block diagram of joint three-dimensional object localization in accordance with the present principles.

FIG. 4 is a block diagram of a moving object localization system in accordance with the present principles.

DETAILED DESCRIPTION

Embodiments of the present invention use monocular structure-from-motion (SFM) to adaptively estimate a ground plane by combining cues from SFM, dense inter-frame stereo, and object detection. Object pose is estimated by intensity alignment, since pose estimation based on very few sparse features is unstable. Dense features are tracked using a one-dimensional optical flow mechanism that imposes epipolar constraints. Outliers in the tracks are eliminated using the estimated object pose. The SFM cues employed by the present embodiments minimize reprojection errors of the tracked three-dimensional points, while an object detection cue fits the best three-dimensional bounding box that is most consistent with a sequence of two-dimensional bounding boxes having detection scores that are highest in relation to the object tracking output.

The present embodiments provide superior accuracy in object localization by using a variety of cues. In addition, the localization is more stable due to the use of dense tracking and intensity alignment for pose computation, and is performed more quickly because motion segmentation and complex inference are not used.

Referring now to FIG. 1, a block diagram of object localization is shown. An input video 102 is used in conventional 2D object detection 104 and adaptive ground plane detection 106. In addition, real-time monocular SFM 108 is used to inform both the adaptive ground plane detection 106 as well as dense 3D estimation 110 The results of the 2D object detection 104, the adaptive ground plane detection 106, and the dense 3D estimation 110 are used in joint 3D object localization 112 to determine a location of an object in three dimensions.

Referring now to FIG. 2, a more detailed view of the dense 3D estimation 110 is shown. Dense 3D estimation 110 uses intensity-aligned object pose computations 202 and epipolar-guided dense tracking 204.

For intensity-aligned object pose computation 202, an object's pose is defined by the vector Ω=(x₀, z₀, ψ, θ, φ, h)^T, where x₀ and z₀ are coordinates of the object's origin in the camera's coordinates, corresponding to the center of the line where the back plane of the object intersects the ground, ψ, θ, and φ are pitch, roll, and yaw angles of the object, and h is the height of the camera from the ground plane. Given an object pose at time t, the pose at time t+1 is determined as the one that best aligns the intensities I at all pixels in the projected image according to:

$\underset{{\Omega }_{t+1}}{\min }\sum _{i=1}^N{\left(I_t\left({\pi }_{{\Omega }_t}\left(x_i\right)\right)-I_{t+1}\left({\pi }_{{\Omega }_{t+1}}\left(x_i\right)\right)\right)}^2$

where π_Ωstands for the projection of a 3D point in object coordinates to the 2D image plane.

Epipolar-guided dense tracking 204 estimates TV-L1 regularized optical flow between frames t and t+1. Flow estimation is performed only within the detection bounding boxes to maximize computational efficiency. The optical flow vectors are constrained to satisfy epipolar geometry. This is used to reduce the optical flow problem from a two-dimensional search to a one-dimensional search. The feature tracks are validated and triangulated. The optical flow problem may be expressed as:

E=∫ _Ω [λ|I ₀(x)−I₁(x+p₀+up)|+∇up]dx

where x is an image location, p₀ is a point on the image, p is the direction of an epipolar line through p₀, and u is the distance along the epipolar line.

Referring now to FIG. 3, additional detail on joint optimization for 3D object localization 112 is shown. A cost function incorporates costs from object bounding boxes, SFM, object detection, and priors from ground plane estimation, trajectory smoothness, and object size. A Levenberg-Marquardt optimization is performed to find the hest 3D hounding box that minimizes the cost function:

ε=ε′_SFM+λ_oε′_obj+λ_pε′_prior

where ε′_SFM is a cost from SFM, ε′_obj is a cost from object bounding boxes, and ε′_prior is a prior from the combined smoothness and object size priors. Exemplary parameter values may be λ₀=0.7 and λ_p=2.7. The quantity ε′_obj is defined as:

ε′_obj=ε′_box+λ_dε′_det

where ε′_box is an object bounding box error, ε′_det is a detection cost, and the parameter λ_d has an exemplary value of 0.03. The quantity ε′_prior is defined as:

ε′_prior=ε′_smooth+λ_sε′_size

where ε′_smooth is a trajectory smoothness prior, ε′_size is an object size prior, and an exemplary value for the parameter is λ_s=0.03.

3D Object localization 112 uses SFM cues 302, bounding box cues 304, and object detection cues 306. Finding object localization cost from SFM cues in block 302 tracks i=1, . . . , N objects in a scene, with j=1, . . . , M features on each object to be tracked in frames t=s_i, . . . , e_i, where s_i and e_i are the first and last frames of the object i. Then, if ū_j^t=(ū_j^t, v_j^t)^T is the observed projection and ū_j^t=(u_j^t, v_j^t) is the matched 2D pixel in frame t, an SFM reprojection error for the feature tracks may be defined as:

$E_{\mathrm{SFM}}=\sum _{i=1}^N\sum _{t=s_i}^{e_i}\sum _{j=1}^M{\left(u_j^t-{\stackrel{\_}{u}}_j^t\right)}^2+{\left(v_j^t-{\stackrel{\_}{v}}_j^t\right)}^2$

where u and v are the coordinates of a 2D pixel on the image plane.

Defining N=[n_α, n_β, n_γ], where n_γ=(−n₁, n₃, −n₂)^T, n_β=−n, and n_α=n_β×n_γ, then given a 3D point {tilde over (x)}₀ in the object coordinate system, the transformation from object to camera coordinates is given by {tilde over (x)}_c=P_πP_ψ{tilde over (x)}₀, with:

$P_{\pi }=\left[\begin{array}{cc}N& c_o\\ 0^T& 1\end{array}\right],P_{\psi }=\left[\begin{array}{cc}{}^{{\left[{\omega }_{\psi }\right]}_{\times }}& 0\\ 0^T& 1\end{array}\right]$

where ω_ψ=(0, ψ, 0)^T and [.]_x is a cross product matrix. The formulation to project a 3D point {tilde over (x)}₀ in the object coordinate system is then given by the homogeneous relation:

λũ=K[I|0]P_πP_ψ{tilde over (x)}₀

where λ is an unknown scale factor and K is a 3×3 intrinsic camera calibration matrix.

Determining an object localization cost from 2D bounding box cues 304A set of tracklets is considered input for the localization, where a tracklet is an object track. The 2D bounding box cost for localization is given by:

$E_{\mathrm{obj}}=\sum _{i=1}^N\sum _{t=s_i}^{e_i}{\left(b_i^t-D_i^t\right)}^2$

where b_i(t) is the 2D projection of the 3D bounding box and D_i(t) is the 2D bounding box from tracking. Given the eight corners of the bounding box, v_i, then their projections q_i are:

λ_i^t{tilde over (q)}_i^t=K[I|0]P_π^tP_ψ^t{tilde over (v)}_i, i=1, . . . , 8

where λ_i^t are the homogeneous scale factors. Then define:

$b_1^t=\underset{i}{\min }p_i^t,b_2^t=\underset{i}{\min }p_i^t,b_3^t=\underset{i}{\min }q_i^t,b_4^t=\underset{i}{\max }q_i^t$

For a 3D bounding box with vertices v₁, . . . , v₈, the corresponding image points are q₁, . . . , q₈, where each q_i is a 2D pixel on the image plane represented by two numbers, (p_i, q_i).

Object localization cost from object detection cues 306 makes the present embodiments not solely dependent on tracking output. Using s(.) as a function to compute the detection score of a bounding box, the detection term is:

$E_{\det }=\sum _{i=1}^N\sum _{t=s_i}^{e_i}{\left(\frac{1}{s\left(b_i^t\right)}\right)}^2$

where b_i^t is the projection of the respective estimated 3D bounding box.

To build the detection score function s(.) without running the detector, the detection score is computed by fitting a model. The detector provides several candidate bounding boxes and their respective scores. The model is a mixture of Gaussians. At each frame, 4×4 full rank covariance matrices Σ_m are estimated centered at μ_m as:

$\underset{A_m,{\mu }_m,{\Sigma }_m}{\min }\sum _{n=1}^N{\left(\sum _{m=1}^MA_m{}^{-\frac{1}{2}{\epsilon }_{\mathrm{mn}}{\Sigma }_m^{-1}{\epsilon }_{\mathrm{mn}}}-{\delta }_n\right)}^2$

where ε_mn=b_n−μ_m, M is the number of objects, and N is the number of candidate bounding boxes, where the dependence on k has been suppressed for the sake of convenience. A_m is the amplitude of the Gaussian, μ_m is the Gaussian's mean, ε_mn is the distance from the mean, Σ_m is the covariance, and δ_n is a score approximated by this sum of Gaussians.

It should be understood that embodiments described herein may be entirely hardware, entirely software or including both hardware and software elements. In a preferred embodiment, the present invention is implemented in hardware and software, which includes but is not limited to firmware, resident software, microcode, etc.

Embodiments may include a computer program product accessible from a computer-usable or computer-readable medium providing program code for use by or in connection with a computer or any instruction execution system. A computer-usable or computer readable medium may include any apparatus that stores, communicates, propagates, or transports the program for use by or in connection with the instruction execution system, apparatus, or device. The medium can be magnetic, optical, electronic, electromagnetic, infrared, or semiconductor system (or apparatus or device) or a propagation medium. The medium may include a computer-readable storage medium such as a semiconductor or solid state memory, magnetic tape, a removable computer diskette, a random access memory (RAM), a read-only memory (ROM), a rigid magnetic disk and an optical disk, etc.

A data processing system suitable for storing and/or executing program code may include at least one processor coupled directly or indirectly to memory elements through a system bus. The memory elements can include local memory employed during actual execution of the program code, bulk storage, and cache memories which provide temporary storage of at least some program code to reduce the number of times code is retrieved from bulk storage during execution. Input/output or I/O devices (including but not limited to keyboards, displays, pointing devices, etc.) may be coupled to the system either directly or through intervening I/O controllers.

Network adapters may also be coupled to the system to enable the data processing system to become coupled to other data processing systems or remote printers or storage devices through intervening private or public networks. Modems, cable modem and Ethernet cards are just a few of the currently available types of network adapters.

Referring now to FIG. 4, an object localization system 400 is shown. A hardware processor 402 accesses video data stored in a memory 404 to provide location information. A 2D detection module 406 uses the processor 402 to perform two-dimensional object detection on the video while the SFM module 410 performs real-time monocular SFM. The ground plane module 408 estimates a ground plane for objects in a frame using the 2D object detection information and the SFM information. Block 412 performs dense 3D estimation using the processor 402. The joint 3D object localization module uses information from the ground plane module 408 and the dense 3D estimation module 412 to provide object location information.

The foregoing is to be understood as being in every respect illustrative and exemplary, but not restrictive, and the scope of the invention disclosed herein is not to be determined from the Detailed Description, but rather from the claims as interpreted according to the full breadth permitted by the patent laws. Additional information is provided in Appendix A and Appendix B to the application. It is to be understood that the embodiments shown and described herein are only illustrative of the principles of the present invention and that those skilled in the art may implement various modifications without departing from the scope and spirit of the invention. Those skilled in the art could implement various other feature combinations without departing from the scope and spirit of the invention.

Claims

1. A method for moving object localization, comprising: estimating a ground plane in a video frame based on a detected object within the video frame and monocular structure-from-motion (SFM) information;computing object pose for objects in the frame based on the SFM information using dense feature tracking; anddetermining a three-dimensional location for the detected object based on the estimated ground plane and the computed object pose.

2. The method of claim 1, wherein determining the three-dimensional location for the detected object is based on SFM cues, bounding box cues, and object detection cues.

3. The method of claim 1, wherein dense feature tracking comprises determining an object pose based on dense intensity alignment and tracking dense features based on an epipolar guided optical flow.

4. The method of claim 1, wherein determining the three-dimensional location comprises a Levenberg-Marquardt optimization that finds a three-dimensional bounding box that minimizes a cost function.

5. The method of claim 4, wherein the cost function is: ε=ε′SFM+λoε′obj+λpε′prior

6. The method of claim 1, wherein motion segmentation and complex inference are not used to speed object localization.

7. A system for moving object localization, comprising: a processor configured to estimate a ground plane in a video frame based on a detected object within the video frame and monocular structure-from-motion (SFM) information, to compute object pose for objects in the frame based on the SFM information using dense feature tracking, and to determine a three-dimensional location for the detected object based on the estimated ground plane and the computed object pose.

8. The system of claim 7, wherein the processor is configured to determine the three-dimensional location for the detected object based on SFM cues, bounding box cues, and object detection cues.

9. The system of claim 7, wherein the processor is configured to track dense features by determining an object pose based on dense intensity alignment and tracking dense features based on an epipolar guided optical flow.

10. The system of claim 7, wherein the processor is configured to determine the three-dimensional location using a Levenberg-Marquardt optimization that finds a three-dimensional bounding box that minimizes a cost function.

11. The system of claim 10, wherein the cost function is: ε=ε′SFM+λoε′obj+λpε′prior

12. The system of claim 7, wherein the processor is configured not to use motion segmentation or complex inference to speed object localization.