Automatic Video Based Perception of Legitimate Persons Using Computer Vision Techniques
Ankit Kumar Sah, Nickolas Savarimuthu · Pattern Recognition and Image Analysis · 2021
Abstract The purpose of any surveillance investigation is to look for recurrent theft, vandalism, missing person searches, fraud, crime, etc. It is a general practice of observing a subject for documenting the whereabouts or interactions of the same. Physical observation though reasonable is hectic, exorbitant, and time consuming. The system must be able to distinguish between legitimates and strangers. Towards the same, an automatic detection system is developed to detect only the registered persons of the media lab. Two variants of YOLO viz., YOLOv3-Tiny, and YOLOv2 were utilized for fast identification, classification, and detection of registered persons on live stream video outputted by the network camera installed at the place. YOLOv2 achieved 96.25% mean average precision whereas YOLOv3-Tiny registered 95.05% with average detection times of 0.06 and 0.04 s per frame respectively. The models outperformed Faster R-CNN detection methodology with regards to mean average precision and detection time utilized for the same task. The achieved results were satisfactory and promising with reference to fast recognition, speed, and correctness.