基于时空双分支网络的行为检测与识别技术研究

Research on behavior detection and recognition technology based on spatiotemporal dual branch network

针对传统行为识别方法难以适应复杂的电厂环境,且未充分利用监控视频的时序信息等问题,提出了一种基于时空双分支网络的行为检测与识别技术。该技术利用时空双分支网络提取图像特征,分别基于卷积神经网络、循环神经网络获取图像的空域及时域特征,且使用混合组卷积与横向连接完成特征融合。同时将融合特征作为Softmax分类函数的输入,并经过分数计算得到行为类型。以某电厂的视频监控数据集为样本进行的实验分析结果表明,所提技术方案的行为识别准确率高达94%,且收敛速度快,优于其他对比技术,能够有效解决电厂的行为检测与识别问题。

Aiming at the problems that the traditional behavior recognition methods are difficult to adapt to the complex power plant environment and do not make full use of the monitoring video timing information,a behavior detection and recognition technology based on spatio⁃temporal dual branch network is proposed in this paper.The spatiotemporal double branch network is used to extract image features,the spatial and temporal features of the image are obtained based on Convolutional Neural Network and Recurrent Neural Network respectively,and the mixed group convolution and transverse connection are used to complete the feature fusion.At the same time,the fusion feature is taken as the input of Softmax classification function,and the behavior type is obtained through score calculation.Taking the video monitoring data set of a power plant as the sample,the experimental analysis results show that the behavior recognition accuracy of the technical scheme proposed in this paper is as high as 94%,and the convergence speed is fast,which is better than other comparison technologies,and can effectively solve the problem of behavior detection and recognition in power plant.