Research On Multi-Carrier Based Abnormal Event Detection

Multi camera surveillance network has been widely laying in key public areas.In recent years,the enthusiasm of the public to participate in social and public activities has been increasing,and the phenomenon of various aggregation activities frequently appears in public areas,followed by the increasingly prominent problem of security risks.To achieve a comprehensive and effective monitoring of public areas,such as railway stations,bus stations,large squares,relevant departments have completed the surveillance camera network in that areas.However,the existing monitoring methods still can not complete the timely response and treatment for the abnormal events occurred in public areas.On the one hand,the monitoring method that directing observation from security person is too low efficiency.With the expansion of the monitoring area,the work intensity of the staff also increases;In addition,a long time continuous observation of the video will cause visual fatigue,resulting in a large number of false alarm and error detection.On the other hand,the existing intelligent video surveillance system is limited.The intelligent video surveillance system can not meet the complex application requirements;In addition,the existing monitoring technology almost for single camera,can not play the advantages of multi camera surveillance network.Therefore,it is an important means to build large public areas joint monitoring system.First,the system quickly determines the suspicious area of abnormal events from whole area.Then,through the multi carrier,system achieve the precise monitoring of the crowd in the suspicious area.Finally,the observation data are pre-processed,and the efficient algorithm is used to realize the rapid judgment of the abnormal events and timely warning.In this dissertation,a number of common problems in the large public areas joint monitoring system are deeply studied.The main research results and contributions are as follows:(1)An abnormal events detection approach is proposed based on the saliency attention mechanism.For the problem of fast location of suspicious events in large scale scene,our approach does not require any training process.By using appearance,contour,velocity and direction of motion cues,and combined with multi-scale analysis technology,the algorithm can realize the location of suspicious events in real-time.(2)A running event detection algorithm based on two-steam convolutional network fusion architecture is proposed to solve the problem of moving object detection based on carrier.Firstly,a simple but efficient spatial-temporal filtering algorithm is used to extract the moving object candidates.Then,a novel two-stream convolutional network fusion architecture is proposed to the running target detection with high accuracy and low false alarm.(3)Focus on the action localization problem in unconstrained video,an action tube extraction algorithm is proposed via spatial actionness estimation and temporal path searching.Algorithm firstly using Faster R-CNN to obtain the action object candidates.Then,Markov Chain Mento Carlo approach is used to link all candidates that belong to same action person in all frames.Through the completion within frames,complete spatio-temporal action sequence is finally obtained.(4)To solve the problem of high accuracy recognition of human action,a human action recognition approach based on sequence tensor decomposition is proposed.Motion feature is the key to action recognition.On the basis of the appearance and short-term motion,our algorithm uses the sequence tensor decomposition algorithm to obtain the global motion features.A multi-stream deep convolutional network fusion architecture is proposed to integrate all features to achieve high accuracy human action recognition.In addition,in order to effectively exploit the long-term temporal dependence,Gated Recurrent Unit(GRU)model is introduced in our architecture to further improves the recognition accuracy.