Research on Pedestrian Intrusion Detection in Static Scenes

Abstract

Security system is an important technical means of implementing security prevention and control, and its use in the field of security technology prevention is becoming more and more widespread in the current situation of expanding demand for security. The security systems used now primarily mainly rely on human visual judgment, which demonstrate the lack of intelligent analysis of video content. Static Pedestrian Intrusion Detection (SPID), which determines whether a pedestrian invades a target area in a static scene, is an important vision task in the field of intelligent video surveillance, and has a wide range of applications in scenarios such as intelligent security. To address the problem of static pedestrian intrusion detection data construction, this paper fully investigates the data set and provides sufficient data preparation for the study of this task. This paper proposes a multi-task deep network model based on target detection region segmentation and fast pedestrian detection to achieve accurate pedestrian intrusion determination in static scenes using the powerful nonlinear feature extraction capability of the network. To solve the real-time problem, the model proposes two mobile network optimization strategies, feature sharing and feature cropping, to reduce the computational complexity of the algorithm. Experimental results show that the proposed model achieves 83.1% accuracy and 20.4 FPS detection speed on the static pedestrian intrusion detection datasets, outperforming existing algorithms in terms of both accuracy and speed to achieve end-to-end real-time pedestrian intrusion detection.