Research On Abnormal Behavior Detection Based On Multiple Instance Learning

A large number of surveillance devices are used in public places to alleviate the public safety problems caused by population growth.Anomalous behaviour detection algorithms,one of the core of intelligent surveillance technologies,are currently a relatively hot topic in the field of video analytics.Weakly supervised correlation algorithms,as the mainstream anomalous behaviour detection algorithm,save annotation costs and facilitate the formation of large datasets because they only require video-level labels instead of frame-level labels,but at the same time the dataset also contains a large number of noisy labels.At the same time,the current algorithms have limitations in capturing temporal contextual information of the video.In order to reduce the interference of noisy labels on classifier training,this thesis proposes a joint training anomaly detection algorithm based on co-regularisation.The algorithm mainly consists of two parallel classifiers simultaneously scoring video snippets in the dataset,followed by a small-loss selection strategy to filter out samples containing clean labels to back-propagate and update the parameters of the network.Experiments on publicly available datasets show that this joint training paradigm outperforms traditional multi-instance ranking algorithms.Since videos have continuous back-and-forth causality,temporal context information plays a very important role in anomaly detection.To capture the temporal contextual relationships over long and short distances,an adaptive multi-scale temporal network is proposed in this paper.The network consists of multiple parallel dilated convolutions to obtain temporal context information at different time scales,and then the neural network assigns weights to multiple branches to obtain adaptive multi-scale temporal context information.Visual analysis of the anomaly scores shows that the model with the inclusion of temporal contextual information has better anomaly detection performance.The optimised model was experimented on the publicly available large anomalous behaviour detection datasets Shanghai Tech,UCF-Crime and XD-Violence.The algorithm in this paper is competitive compared to the current state-of-the-art algorithms.Notably,the algorithm achieves the best current classification performance on the Shanghai Tech dataset and the XD-Violence dataset.In particular,the Shanghai Tech dataset achieved an AUC of 97.81% and the XD-Violence dataset achieved an AP of 80.91%.