Pedestrian Detection and Behaviour Characterization for Video Surveillance Systems

Nowadays, to identify potentially dangerous situations created by pedestrians, performing video surveillance systems are more than necessary. The application presented in this paper is designed to automatically detect individual pedestrians and analyze and characterize their behavior. Based on this result, an informed security decision can be made. The implementation uses YOLO (You-Only-Look-Once) [I] to detect people in a real-time video stream and evaluates and characterizes their movement, based on data collected from the current video frame and a certain number of past video frames. A series of tests revealed that our system successfully detects both regular and irregular pedestrian trajectory type and also the type of movement: walk, run or stand. The processing speed of up to 77 FPS for YOLOv3-tiny and 21 FPS for YOLOv3 qualifies our solution for real-time operation in a surveillance system, if it is running on a computer platform equipped with a NVIDIA GPU with CUDA capabilities.