Foreground Object Detection and Regression-based Video Anomaly Detection Method
Video anomaly detection finds wide applications in the field of intelligent security.Methods based on generative models have garnered extensive attention in academia due to their powerful generative capabilities.However,such methods typically involve a large number of parameters and often rely on a vast amount of training data,limiting their applicability in real-world scenarios.This paper proposes a video anomaly detection method based on foreground object detection and regression(FODR-VAD).Firstly,foreground objects are detected using an object detector,and spatiotemporal cubes centered around these objects are constructed.Secondly,pseudo-anomalous data is created using a random shuffling approach.Finally,the video anomaly detection problem is transformed into a regression task,optimizing feature representation under the supervised learning paradigm.With the model parameter count less than 1 million and using less than half of the training set,the proposed method achieves Micro-AUC scores of 99.09%,88.16%,and 78.47%on the UCSD Ped2,CUHK Avenue,and ShanghaiTech datasets,respectively.The results demonstrate that the proposed method significantly reduces the requirement for training data while ensuring high anomaly detection capability.
video anomaly detectionpseudo anomalysupervised learningregressionspatio-temporal cube
万方数据
维普
