Surveillance Video Event Detection In Complex Scence

Intelligent video surveillance system has been paid more and more attention by the public in recent years. It plays an important role in the field of public security. Event detection is the core of the intelligent surveillance video system. The main technologies involved include target recognition and detection, target tracking, behavior analysis, event recognition and detection.In this thesis, we focus on pedestrian detection and event detection in complex scenes.An effective head-shoulder detection algorithm for pedestrian detection based on Faster R-CNN framework is proposed. The part-based scheme effectively suppresses the appearance variations of pedestrians caused by heavy occlusion. The proposed skip-pooling, hard negative mining and bounding box voting strategies make the model able to capture discriminative information in pedestrian head-shoulders. Experiments on the SED-PD show that the performance of pedestrian detection algorithm proposed in this paper is the best in crowded surveillance videos, which surpasses other excellent pedestrian detection algorithms by a large margin.In Caltech pedestrian datasets, the strategies we propose to improve Faster R-CNN are proved to be effective. Competitive performance is achieved among other state of art detection methods. More importantly, our method is comparable in speed to other methods.As for event detection, method based on Convolutional Neural Network and key-pose is utilized. This method has ranked first and second respectively in the ObjectPut and PersonRuns event detections of the TRECVID-SED international evaluation in 2015. Moreover, the improved algorithm of PersonRuns event detection has got first place in TRECVID-SED 2016.