Surveillance Video Event Detection In Complex Scence

Intelligent video surveillance system is becoming more and more important and its key technology is event detection.In this thesis we focus on three aspects:pedestrian detection,event detection and multi-GPU parallel optimization.An effective CNN-based head-shoulder detecting method is proposed.The Head-shoulder Cascade Deep Network(HsNet)is the main model.In this model,the part-based scheme effectively restrains the appearance variations of pedestrians caused by heavy occlusion.The deep network captures discriminative information of visible body parts.In addition,the cascade architecture enables very fast detection.To encourage the research of pedestrian detection,we manually label part of the surveillance video dataset,i.e.,TRECVID SED 2008.We get a total of 404,000 pedestrian samples and name it as TRECVID SED Pedestrian Dataset(SED-PD).The experiments on SED-PD show that our method achieves very competitive performance compared with state-of-the-art methods in crowded surveillance videos.More importantly,our method is significantly faster.An event detection method based on CNN and key-pose is presented.With this method,we got 1th and 2nd places respectively in Embrace and Pointing detections of TRECVID SED 2014.And we rank both 1th in Embrace and Pointing detections of TRECVID SED 2015 with the improved version of this method.Further more,we also explore how to fuse the spatial and temporal informations of video events with CNN.In the multi-GPU parallel optimization,the Asynchronous Stochastic Gradient Descent(ASGD)Algorithm implemented in this thesis is 1.920,1.724 and 1.281 times faster(2GPU)than Stochastic Gradient Descent(SGD)Algorithm(1GPU)respectively in the training of LeNet,CifarNet and AlexNet models.And its overall performance is higher than another multi-GPU parallel algorithm(HogWild).