Key Techniques Research Of Surveillance Video Feature Extraction Based On Stream Computing

With the arrival of big data generation, vast amount of video data have been collected, transmitted, distributed and Widely applied all over the world. A serious expansion of information caused by the characteristics of huge amount and low level abstraction of the video data. So how to effectively describe the video information and achieve efficient content-based video retrieval turns to be a hot topic. Video feature is a description of the video content. The video feature extraction feature can establish an index for the massive videos. So it provides the foundation for users to quickly retrieve the video segments from the mass interested video, and effective support for much business information of public security, transportation and other industries.In this paper, the author propose a real-time based method for extracting real-time surveillance video features according to the characteristics of surveillance video and the real-time demand for the surveillance scene. All works are done as follows:1.Putting forward a suitable method for surveillance video feature extraction. According to the characteristics of the fixed feature of monitoring camera, instead of the shot based video segmentation method, setting the SIFT matching as the judgement standard of the similarities between frames, choosing a kind of similar content based clustering approach for surveillance video segmentation.2.Putting forward a kind of key frame extraction method based on the number of SIFT feature points. In this method, the number of feature points is regarded as the key frame selection standard. As for a sequence of frames with similar content, selecting frames with the most feature points as the video key frames.3.Putting forward a parallel SIFT feature extraction method based on the data distribution strategy. Depending on the parallel computing ability of multiprocessor cluster, the method divide the frame into multiple data block, and send the block to multiple nodes for parallel processing. By accelerating the single frame of SIFT feature extraction process, the method improve the efficiency of the video features extraction.4.For the real-time demand in monitoring scenarios, putting forward a kind of surveillance video feature extraction scheme based on Storm. Achieving the extraction effect of monitoring videos' real-time features.The experimental results show that the clustering based on content similar frames of surveillance video feature extraction method can effectively split the contents of similar video clips from surveillance video. The key frame extraction method based on maximum feature point principle can effectively extract key frames with low redundancy, and achieve good results of video feature extraction. The data distribution based parallel SIFT feature extraction method effectively improves the SIFT feature extraction speed, and the surveillance video feature extraction scheme based on storm can achieve good real-time performance and scalability.