UMTSS: a unifocal motion tracking surveillance system for multi-object tracking in videos.

Multiple object detection and tracking play a very crucial role in solving several elementary problems in real-time surveillance video analysis and computer vision. However, it is a challenging problem because real-time surveillance videos are typically affected by a variety of adverse environmental effects. In this work, we propose a novel surveillance framework, called a unifocal motion tracking surveillance system (UMTSS), for multi-object tracking in real-time videos. The proposed UMTSS combines two significant steps. First, a Faster-RCNN with inception-v2 model is employed here to detect multi-objects efficiently in each video frame. Then, a unifocal feature-based KLT (Kanade-Lucas-Tomasi) method is proposed for tracking objects across the video frames based on region proposals generated by the object detector in the previous phase. Also, we have proposed a new tracking parameter, called dynamic tracking accuracy (DTA), to quantify the performance of the tracking algorithms. The performance of our UMTSS has been evaluated on five standard crowd video databases, namely CrowdHuman, PETS, UCSD, AGORASET and CRCV, and compared with state-of-the-art methods in terms of different qualitative and quantitative measures. It has been observed that our UMTSS outperforms the state-of-the-art methods. [ABSTRACT FROM AUTHOR]