Moving object localization in 3D using a single camera
View Patent ↗Systems and methods are disclosed for autonomous driving with only a single camera by moving object localization in 3D with a real-time framework that harnesses object detection and monocular structure from motion (SFM) through the ground plane estimation; tracking feature points on moving cars a real-time framework to and use the feature points for 3D orientation estimation; and correcting scale drift with ground plane estimation that combines cues from sparse features and dense stereo visual data.
1. A computer vision method for autonomous driving that uses only a single camera, comprising:
moving object localization in 3D with a real-time framework that harnesses object detection and monocular structure from motion (SFM) through a ground plane estimation;
tracking feature points on moving cars a real-time framework to and use the feature points for 3D orientation estimation;
correcting scale drift with ground plane estimation that combines cues from sparse features and dense stereo visual data; and
applying a Kalman filter framework that adapts the fusion observation covariances at every frame to reflect the relative uncertainty of each cue, the Kalman filter comprising a model of state of evolution as
x k =Ax k-1 +w k-1 ,p ( w ): N (0 ,Q ),
z k =Hx k-1 +v k-1 ,p ( v ): N (0 ,U ),
where x represents a state variable, z an observation, Q and U represent covariances of process and observation noise and are assumed to be zero mean multivariate normal distributions as indicated by p(w):N(0,Q) and p(v):N(0,U), A represents state transition, w a process noise, H an observation that maps true state space into observed state space, and v an observation noise.
2. The method of claim 1 , comprising incorporate cues from multiple methods of ground plane estimation.
3. The method of claim 1 , comprising applying a framework that accounts for per-frame relative confidences using models learned from extensive training data.
4. The method of claim 1 , comprising training with frames from a dataset, wherein models are learned that relate an observation covariance for each cue to error behaviors of underlying variables.
5. The method of claim 1 , comprising performing adaptive estimation of observation covariances for cue combination.
6. The method of claim 1 , comprising localization framework that combines information from object bounding boxes and SFM feature tracks, through the ground plane.
7. The method of claim 1 , comprising combining SFM and object bounding boxes through the adaptive ground plane for 3D localization of near and distant objects.
8. The method of claim 1 , comprising performing epipolar update with a rough vanishing point estimation.
9. A computer vision system for autonomous driving that uses only a single camera, comprising:
a real-time framework for moving object localization in 3D that harnesses object detection and monocular structure from motion (SFM) through the ground plane estimation,
a real-time framework to track feature points on moving cars and use them for 3D orientation estimation,
computer code to correct scale drift using ground plane estimation that combines cues from sparse features and dense stereo visual data, and
a Kalman filter framework that adapts the fusion observation covariances at every frame to reflect the relative uncertainty of each cue, the Kalman filter comprising a model of state of evolution as
x k =Ax k-1 +w k-1 ,p ( w ): N (0 ,Q ),
z k =Hx k-1 +v k-1 ,p ( v ): N (0 ,U ),
where x represents a state variable, z an observation, Q and U represent covariances of process and observation noise and are assumed to be zero mean multivariate normal distributions as indicated by p(w):N(0,Q) and p(v):N(0,U), A represents state transition, w a process noise, H an observation that maps true state space into observed state space, and v an observation noise;
wherein the system includes a processor for processing the real time framework and the computer code.
10. The system of claim 9 , wherein the framework incorporates cues from multiple methods of ground plane estimation.
11. The system of claim 9 , wherein the framework that accounts for per-frame relative confidences, using models learned from extensive training data.
12. The system of claim 9 , wherein models are learned that relate an observation covariance for each cue to error behaviors of underlying variables.
13. The system of claim 9 , comprising an adaptive estimator of observation covariances for cue combination.
14. The system of claim 9 , comprising localization framework that combines information from object bounding boxes and SFM feature tracks, through the ground plane.
15. The system of claim 9 , wherein SFM and object bounding boxes through the adaptive ground plane for 3D localization of near and distant objects.
16. The system of claim 9 , comprising an epipolar update with a rough vanishing point estimation.