Methods and system for automated in-field hierarchical training of a vehicle detection system
View Patent ↗Disclosed are methods and systems for detecting one or more vehicles in video captured from a deployed video camera directed at a parking region. According to one exemplary embodiment, disclosed is a method of training a deployed classifier associated with the video camera, where a generic classifier is initially used to obtain high confidence training samples from the video camera, the high confidence training samples subsequently used to train the deployed classifier.
1. A computer implemented method of detecting a vehicle in video, the video acquired from a fixed parking occupancy video camera including a field of view associated with a parking region, the method comprising:
a) capturing video from the fixed parking occupancy video camera; and
b) performing a sliding window-based space search for a vehicle within the captured video, the sliding window-based space search extracting one or more features associated with each of a plurality of windows and accessing an operatively associated second classifier to classify each window as including a vehicle or not including a vehicle,
wherein the second classifier is trained using high confidence training samples acquired from samples of video obtained from the fixed parking occupancy video camera operatively associated with a first classifier, the samples including positive (vehicle) samples of video and negative (non-vehicle) samples of video, and
wherein the first classifier is trained with video captured from another fixed parking occupancy video camera including a different field of view associated with another parking region.
2. The method of detecting a vehicle in video according to claim 1 , wherein the first classifier is configured as a generic classifier.
3. The method of detecting a vehicle in video according to claim 1 , wherein the second classifier is trained using high confidence positive (vehicle) training samples acquired from the fixed parking occupancy video camera operatively associated with the first classifier.
4. The method of detecting a vehicle in video according to claim 1 , wherein the second classifier is trained using high confidence negative (non-vehicle) training samples acquired from the fixed parking occupancy video camera without the use of the first classifier.
5. The method of detecting a vehicle in video according to claim 1 , wherein the high confidence positive training samples are obtained by configuring a vehicle detection threshold to minimize false positive detections of vehicle associated with the samples of video.
6. The method of detecting a vehicle in video according to claim 1 , wherein the high confidence negative training samples are obtained by configuring a vehicle detection threshold to minimize false negative detections associated with the samples of video.
7. The method of detecting a vehicle in video according to claim 1 , wherein the first classifier is trained with manually labeled image data captured from another fixed parking occupancy video camera including a different field of view associated with another parking region.
8. The method of detecting a vehicle in video according to claim 1 , wherein the second classifier is trained using a machine learning process.
9. The method of detecting a vehicle in video according to claim 1 , wherein the second classifier is trained by iteratively modifying all or part of the first classifier to include a current version of the second classifier, acquiring high confidence training samples from samples of video obtained from the fixed parking occupancy video camera operatively associated with the first classifier, and training the second classifier using the high confidence training samples.
10. A computer implemented method of detecting a video, the video acquired from a fixed parking occupancy video camera including a field of view associated with a parking region, the method comprising:
a) capturing video from the fixed parking occupancy video camera; and
b) performing a sliding window-based space search for a vehicle within the captured video, the sliding window-based space search extracting one or more features associated with each of a plurality of windows and accessing an operatively associated second classifier to classify each window as including a vehicle or not including a vehicle,
wherein the second classifier is trained using high confidence training samples acquired from samples of video obtained from the fixed parking occupancy video camera operatively associated with a first classifier, the samples including positive (vehicle) samples of video and negative (non-vehicle) samples of video, and
wherein the positive (vehicle) samples of video are associated with a sub-region of a video-frame, the sub-region related to a ROI (Region of Interest).
11. The method of detecting a vehicle in video according to claim 10 , wherein the generic classifier is configured as one or more of a weak classifier, a color-based vehicle detector, and a multi-view vehicle classifier.
12. A vehicle detection system associated with a parking region, the vehicle detection system comprising:
a fixed parking occupancy video camera including a field of view associated with the parking region; and
a controller operatively associated with the fixed parking occupancy video camera, the controller configured to execute computer instructions to perform a process of detecting a vehicle in video including:
a) capturing video from the fixed parking occupancy video camera; and
b) performing a sliding window-based space search for a vehicle within the captured video, the sliding window-based space search extracting one or more features associated with each of a plurality of windows and accessing a second classifier to classify each window as including a vehicle or not including a vehicle,
wherein the second classifier is trained using high confidence training samples acquired from samples of video obtained from the parking occupancy video camera operatively associated with a first classifier, the samples including positive (vehicle) samples of video and negative (non-vehicle) samples of video, and
wherein the first classifier is trained with video captured from another fixed parking occupancy video camera including a different field of view associated with another parking region.
13. The vehicle detection system associated with a parking region according to claim 12 , wherein the generic classifier is configured as one or more of a weak classifier, a color-based vehicle detector, and a multi-view vehicle classifier.
14. The vehicle detection system associated with a parking region according to claim 12 , wherein the first classifier is configured as a generic classifier.
15. The vehicle detection system associated with a parking region according to claim 12 , wherein the second classifier is trained using high confidence positive (vehicle) training samples acquired from the fixed parking occupancy video camera operatively associated with the first classifier.
16. The vehicle detection system associated with a parking region according to claim 12 , wherein the second classifier is trained using high confidence negative (non-vehicle) training samples acquired from the fixed parking occupancy video camera without the use of the first classifier.
17. The vehicle detection system associated with a parking region according to claim 12 , wherein the high confidence positive training samples are obtained by configuring a vehicle detection threshold to minimize false positive detections of vehicle associated with the samples of video.
18. The vehicle detection system associated with a parking region according to claim 12 , wherein the high confidence negative training samples are obtained by configuring a vehicle detection threshold to minimize false negative detections associated with the samples of video.
19. The vehicle detection system associated with a parking region according to claim 12 , wherein the first classifier is trained with manually labeled image data captured from another fixed parking occupancy video camera including a different field of view associated with another parking region.
20. The vehicle detection system associated with a parking region according to claim 12 , wherein the second classifier is trained using a machine learning process.
21. The vehicle detection system associated with a parking region according to claim 12 , wherein the second classifier is trained by iteratively modifying all or part of the first classifier to include a current version of the second classifier, acquiring high confidence training samples from samples of video obtained from the fixed parking occupancy video camera operatively associated with the first classifier, and training the second classifier using the high confidence training samples.
22. A vehicle detection system associated with a parking region, the vehicle detection system comprising:
a fixed parking occupancy video camera including a field of view associated with the parking region; and
a controller operatively associated with the fixed parking occupancy video camera, the controller configured to execute computer instructions to perform a process of detecting a vehicle in video including:
a) capturing video from the fixed parking occupancy video camera; and
b) performing a sliding window-based space search for a vehicle within the captured video, the sliding window-based space search extracting one or more features associated with each of a plurality of windows and accessing a second classifier to classify each window as including a vehicle or not including a vehicle,
wherein the second classifier is trained using high confidence training samples acquired from samples of video obtained from the parking occupancy video camera operatively associated with a first classifier, the samples including positive (vehicle) samples of video and negative (non-vehicle) samples of video, and
wherein the positive (vehicle) samples of video are associated with a sub-region of a video frame, the sub-region related to a ROI (Region of Interest).