Comparing state-of-the-art methods of detection and tracking people on security cameras video

The use of security cameras allows to perform video surveillance. Nevertheless, little has been done to get other type of information from the videos. This project develops a people counting system for video from detection and tracking algorithms developed on Python 3. The use of trackers and detectors makes the system more robust to avoid overcounting problems. Comparisons were made to evaluate three different detection algorithms: YOLO, SSD and OpenPose; and six tracking algorithms developed on OpenCV: Boosting, MIL, KCF, TLD, Median Flow, MOSSE y CSRT, in order to implement the best algorithms. The system manages to obtain a relative error of maximum counting of 40%, having trouble with sub counting of people in the video.