Camera assisted tracking of objects in a radar system
Camera-assisted tracking of point objects in a radar system is provided. An extended Kalman filter framework based on both radar and camera observations is used to track point objects detected in frames of radar signal data. This framework provides a minimum mean square estimation of the current state of a point object based on previous and current observations from both frames of radar signals and corresponding camera images.
1. A method comprising:
detecting a point object in a frame of radar signals;
estimating a three dimensional (3D) location of the detected point object;
receiving an image captured by a camera;
determining whether the detected point object is within view of the camera based on the received image and the estimated 3D location of the detected point object;
estimating a two dimensional (2D) pixel location of the detected point object in the image based on the received image, the estimated 3D location of the detected point object, and a camera-radar calibration matrix; and
when the detected point object is within the view of the camera, updating a tracking vector and an error covariance matrix for the detected point object using a camera-only observation update computation based on the estimated 2D pixel location of the detected point object in the image.
2. The method of claim 1 , wherein estimating the 3D location of the detected point object comprises:
receiving spherical coordinates of a detected point object, a range rate of the detected point object, a first variance for the received spherical coordinates, and a second variance for the received range rate; and
estimating the 3D location of the detected point object based on the received spherical coordinates, received, range rate, received first variance, and received second variance.
3. The method of claim 1 , further comprising:
performing a time update of the tracking vector and the error covariance matrix in the tracking data of the detected point object by computing an a priori process state vector and an a priori error covariance matrix for the detected point object.
4. The method of claim 3 , wherein determining whether the detected point object is within view of the camera comprises:
mapping a 3D location in the a priori process state vector to a 2D pixel location in the image; and
searching for a pattern block in the image that includes the detected object point by starting the search with a first block centered on the 2D pixel location and subsequently searching for a second block that best matches a previous pattern block from a previous image captured at a previous point in time, wherein the detected point object was included in the previous pattern block.
5. The method of claim 4 , wherein mapping a 3D location comprises using the camera-radar calibration matrix to map coordinates of the 3D location in the a priori process state vector to coordinates of the 2D pixel location in the image.
6. The method of claim 1 , wherein the camera-radar calibration matrix is predetermined.
7. The method of claim 3 , wherein updating tracking data comprises:
performing camera-radar joint observation update computations to update the tracking vector and the error covariance matrix, wherein the estimated 2D pixel location is a center of the block in the image that includes the detected object point and the estimated 3D location is included in the a priori process state vector.
8. The method of claim 1 , further comprising:
updating the tracking vector for the detected point object when the detected point object is not within view of the camera, the updating based on the estimated 3D location of the detected point object.
9. The method of claim 1 , further comprising: performing null observation update computations to update the tracking vector and the error covariance matrix of the detected point object when the detected point object is not within view of the camera.
10. A radar system, comprising:
a processing unit; and
a memory storing instructions that, when executed by the processing unit, cause the radar system to:
detecting a point object in a frame of radar signals;
estimate a three dimensional (3D) location of the detected point object;
receive an image captured by a camera;
determine whether the detected point object is within view of the camera based on the received image and the estimated 3D location of the detected point object;
estimating a two dimensional (2D) pixel location of the detected point object in the image based on the received image, the estimated 3D location of the detected point object, and a camera-radar calibration matrix; and
when the detected point object is within view of the camera, update a tracking vector and an error covariance matrix for the detected point object using a camera-only observation update computation based on the estimated 2D pixel location of the detected point object in the image.
11. The radar system of claim 10 , wherein estimating the 3D location of the detected point object comprises:
receiving spherical coordinates of a detected point object, a range rate of the detected point object, a first variance for the received spherical coordinates, and a second variance for the received range rate; and
estimating the 3D location of the detected point object based on the received spherical coordinates, received, range rate, received first variance, and received second variance.
12. The radar system of claim 10 , wherein the instructions further cause the radar system to:
perform a time update of the tracking vector and the error covariance matrix in the tracking data of the detected point object by computing an a priori process state vector and an a priori error covariance matrix for the detected point object.
13. The radar system of claim 12 , wherein the instructions that cause the radar system to determine whether or not the detected point object is within view of the camera include instructions that cause the radar system to:
map a 3D location in the a priori process state vector to a 2D pixel location in the image; and
search for a pattern block in the image that includes the detected object point by starting the search with a first block centered on the 2D pixel location and subsequently searching for a second block that best matches a previous pattern block from a previous image captured at a previous point in time, wherein the detected point object was included in the previous pattern block.
14. The radar system of claim 13 , wherein the instructions that cause the radar system to map a 3D location include instructions that cause the radar system to use the camera-radar calibration matrix to map coordinates of the 3D location in the a priori process state vector to coordinates of the 2D pixel location in the image.
15. The radar system of claim 10 , wherein the camera-radar calibration matrix is predetermined.
16. The radar system of claim 12 , wherein the instructions that cause the radar system to update the tracking data include instructions that cause the radar system to:
perform camera-radar joint observation update computations to update the tracking vector and the error covariance matrix, wherein the estimated 2D pixel location is a center of the block in the image that includes the detected object point and the estimated 3D location is included in the a priori process state vector.
17. The radar system of claim 10 , wherein the instructions further cause the radar system to:
update the tracking vector for the detected point object when the detected point object is not within view of the camera, the update based on the estimated 3D location of the detected point object.
18. The radar system of claim 10 , wherein the instructions further cause the radar system to:
perform null observation update computations to update a tracking vector and an error covariance matrix of the detected point object when the detected point object is not within view of the camera.