Video_detect.py, Video_process.py, Performance.py, and Video.py contain the functions for video processing.
Faster_RCNN_predict.py is a modified version of the predict function, adapted to save results in the Yolov5 format.
Motion_feature_map_extract.py is the function used to extract the motion vector matrix from the video.
Group_select.py and Manager.py are components of the algorithm.
Running Execution_setting.py will generate the algorithm's results and output them to the output_directory.
| Camera | Length (s) | Description |
|---|---|---|
| Mobile1 | 9651 | Daytime drive in Seattle streets. |
| Mobile2 | 5968 | Drive around Kuwait City. |
| Mobile3 | 2157 | Daytime drive through downtown Vancouver. |
| Mobile4 | 5064 | Drive through Los Angeles downtown. |
| Mobile5 | 2961 | Drive through Chicago downtown. |
| Mobile6 | 297 | Vehicle cameras in different scenarios. |
| Fixed1 | 840 | Relaxed highway traffic near French Alps. |
| Fixed2 | 306 | Urban traffic for detection and tracking. |
| Fixed3 | 2048 | Highway traffic for object recognition. |
| Fixed4 | 904 | Same intersection from different angles. |
Note: Fixed 4 and Mobile 6 have multiple camera sources, 8 and 9 respectively.
The folder motivation_source_video contains the videos of Fixed4 and Mobile6 used in the study, while the folder source_video_demo contains 5-second slices of the Mobile1-5 and Fixed1-3 videos.
The Combined_motion_features matrix is too large, so only a 5-second segment is saved, and it has been processed through a 2D CNN for feature extraction.