PERSON TRACKING IN LIVE VIDEO FEEDS

Abstract

The idea is to utilize the combination of Deep Learning Object Detection, using YOLO (You Only Look Once) and Deep SORT (Simple Online and Realtime Tracking). We are looking into first classifying the person inside the video frames by passing the frames into the neural networks and based on the output co ordinates we track the person using SORT. The idea is to track the person even when his/her face is not visible. There will be utilization of feature extraction and mapping along with frames history. The analysis is fully done in videos after decomposition of videos. The feature like clothes color, location coordinates will be utilized for this purpose. The system will fully be capable of tracking multiple suspects in real-time. The system need not to be fed with facial data, though it will be a useful and implementable for improved accuracy. The system requires manual feeding of target, which can be done by manual targeting or a presumed database with suspicious person data, where descriptors of the person is gives. The system utilizes Kalman Filter of original SORT algorithm for this purpose and the outcome of this system is a warning or analysis on a person.