| Statistic-based object classification for surveillance video has important application value in many fields such as public safety management,traffic dispatching,and construction of large public places in smart cities.In recent years,the use of deep learning models for joint surveillance video content perception in different application scenarios has attracted widespread attention from domestic and foreign researchers.However,surveillance video data has the characteristics of huge volume,various data types,and low-value density.The use of existing high-precision deep learning models has the problems of large calculation volume and slow inference speed,and efficient video content perception cannot be performed.When deploying a deep learning model,it is necessary to manually weigh accuracy and efficiency in combination with actual application scenarios.This paper conducts adaptive inference optimization for the surveillance video object statistical classification application.Firstly,for the optimization space of the surveillance video frame level,reducing the image resolution will improve the video processing efficiency of the deep learning model,which may lead to the loss of object statistical accuracy,but due to the existence of the object statistical classification interval,even if the object statistical results are biased,the final classification accuracy may not decrease.Secondly,for the optimization space of surveillance video segment level,due to the time redundancy of surveillance video,the inference configuration can be adjusted adaptively by sensing the changes of video content,which can improve the efficiency of overall inference and minimize the loss of accuracy.Based on the above two observations,this paper proposes a video analytics inference optimization framework based on statistic-based object classification,namely Swift.In the optimization of the surveillance video frame level,a mechanism for adaptively adjusting the classification threshold is proposed to perceive whether the current inference configuration is reasonable.In the optimization of the surveillance video segment level,a risk decision threshold knob is proposed to adjust the model inference configuration to meet the trade-off between accuracy and efficiency of different applications.Finally,this article conducts a detailed experimental analysis on the Swift framework,which verifies the validity and feasibility. |
