Research On Efficient Processing Mechanisms For Massive Surveillance Videos

High-resolution surveillance cameras have been widely deployed throughout the city,greatly facilitating city management and ensuring social stability.Although these surveillance cameras use video coding standards to dramatically compress their data volume,the huge data volume still puts tremendous pressure on the existing storage system.In addition,in recent years,deep learning-based automatic video analysis technology has developed rapidly.However,complex network model structure has led to high computational cost,making the processing speed of the system insufficient to support realtime analysis for massive surveillance videos.Therefore,efficient processing for massive surveillance videos has important research value and application prospects.Surveillance cameras are usually installed in a specific position and maintain a fixed shooting angle to monitor target objects or abnormal events in the scene.Due to the low occurrence frequency of target events and the repeated appearance of target objects in consecutive video frames,there is a large amount of irrelevant or redundant information in video streams.Frame-by-frame analysis and storage are inefficient,which wastes a lot of computing and storage resources.Therefore,using advanced object detection models and studying the optimized architecture and mechanism of surveillance video processing systems to quickly filter out the irrelevant and redundant information in the videos can greatly reduce computing and storage costs while encuring the prediction accuracy.The main research contents and contributions of this thesis are as follows:Aiming at the video frames irrelevant to the target event,a pipelined multi-stage filtering mechanism is proposed.Using light-weight models to quickly filter out a large number of irrelevant video frames can greatly reduce the workloads of the full-feature model,thereby improving the overall analyzing efficiency of the system.Therefore,a pipelined multi-stage filtering mechanism is proposed,named FFS-VA.FFS-VA first designs a specialized difference detector(SDD)and a specialized network model(SNM)for each video stream,which are usesd to filter out background frames and the frames without target objects respectively.Then,a general light-weight object detection model is used to filter out frames whose target objects are fewer than a predefined threshold.Finally,only the surviving frames are fed into the full-feature object detection model for accurate analysis and detection.FFS-VA designs a global feedback-queue method to alleviate resource competition between different filters at runtime.FFS-VA designs a dynamic batch method to realize the automatic trade-off between throughput and latency.FFS-VA also designs a multi-GPU parallel scheduling mechanism.Experimental results show that FFS-VA can improve the online detection speed by 15 times with an accuracy loss of less than 2%.Aiming at the video segments that do not contain the target event,a joint scheduling mechanism based on light-heavy dual models is proposed.By analyzing the prediction results of general light-weight models and full-feature models,it is found that some general light-weight models can also be used to identify the video segments with target events,so a joint scheduling mechanism based on light-heavy dual models is proposed,named VScan.Among them,the light-weight object detection model is responsible for quickly judging whether video segments contain the target event.Only the potential video segments containing the target event are redirected to the full-feature model for accurate analysis and detection.VScan designs a model selection method and a parameter selection method to ensure the prediction accuracy of the system.VScan designs automatic sampling technology,stream scheduling method,and GPU scheduling strategy to improve hardware efficiency.Experimental results show that VScan significantly boosts the video analyzing throughput by up to 15 times without target event loss.Aiming at the efficient extraction demand for key content in video streams,a three-stage gradual refining mechanism for surveillance video is proposed.Existing surveillance videos generally are addressed by standard video coding technology,but for automatic video analysis,these compressed video streams still have a lot of content redundancy.To achieve efficient storage and rapid analysis,a three-stage gradual refining mechanism for surveillance video is proposed,named VRefine.With the aging of surveillance videos,VRefine uses three consecutive stages to gradually refine the key content: extracting key frames from the raw encoded video streams;condensing the key frames based on the movement vectors;using automatic video analysis to obtain object semantic information.Experimental results show that,VRefine can reduce 42.3%-94.3%storage cost and shorten 46.5%-95.8% analyzing time.The above-mentioned research on efficient processing mechanisms for video frames,video segments,and video streams has greatly improved the analyzing speed and storage efficiency of large-scale surveillance video processing systems,which provides key technical support for the construction of smart cities in the future.