边缘环境下的自适应视频分析

Adaptive video analysis in edge computing

为解决借助深度神经网络对大量交通视频进行分析需要较高计算资源和计算时延较长的问题,介绍了一种基于马尔可夫链的自适应视频分析方法(MCA)。首先,收集交通路况历史视频数据,平均划分为很多个时间区间,对每个时间区间进行汽车数量和平均行驶速度的测量,并基于马尔可夫链模型计算状态转移概率矩阵;其次,调整视频帧采样率获得每个速度状态下的最优视频分析配置;最后,依据当前的速度状态和状态转移概率矩阵预测未来时间区间的速度状态,并选择相应的最优视频分析配置。随机帧采样(SR)准确率较低,每三帧采样一帧(SE3)与MCA准确率均高于90%,但MCA与SE3相比仅需要37.47%的计算资源和37.05%的计算时延。实验结果表明,MCA在对交通视频进行分析时,能有效降低计算资源和计算时延。

Since analyzing a large number of traffic videos using deep neural network requires much computing resource and causes high computational delay,Markov Chain-based adaptive video Analysis method(MCA)was proposed.Firstly,the video data about road traffic was collected and divided into many time intervals.The vehicle number and speed were measured for each time interval,and the state transition probability matrix was calculated based on the Markov chain model.Secondly,the frame sampling rate for the video data was adjusted to obtain the optimal configuration under each speed state.Finally,the speed state of the future time interval was predicted according to the current speed state and the state transition probability matrix,and the corresponding optimal video analysis configuration was selected.Compared with Sampling Randomly(SR)method and Sampling a frame Every 3 frames(SE3)method,the former has low analysis accuracy,MCA only requires 37.47%computing resource and costs 37.05%computing delay of the latter with the average analysis accuracy higher than 90%.The experimental results show that the MCA method can effectively reduce computing resource usage and computational delay when analyzing traffic videos.