面向变电站视频监控终端的目标检测方法及其优化

An Object Detection Method and Its Optimization for Substation Video Surveillance Terminals

变电站视频监控目标检测中,海量高清视频数据传输会对网络带宽、延迟等提出极高的要求,而传统处理方法在应用中存在诸多难以克服的问题,为此提出了一种面向边缘端优化的深度卷积神经网络目标检测方法。首先在粗粒度检测方面利用在线困难实例挖掘(online hard example mining,OHEM)方法和损失函数改进优化模型性能,并借助标签平滑方法防止过拟合;其次采用主干网络替换和剪枝方法实现模型压缩,保持算法推理实时性;最后从细粒度分类方面对所采用的SmallerVGGNet进行网络裁剪修改和多标签分类,确保算法在低功耗前端的轻量化运行。实验结果表明:该方法在粗粒度检测方面相比较于传统算法性能优越,推理速度达到了前端应用的实时性要求,在细粒度分类上也达到了变电站场景下具备不同属性目标的准确分类要求。

In substation video surveillance object detection,mass high-definition(HD)video data transmission may present very high demands for network bandwidth and delay,but there are many invincible problems in traditional treatment methods.Therefore this paper proposes an object detection method based on deeply convolutional neural network(CNN)for edge terminal optimization.For coarse-grained detection,it firstly uses the online hard example mining(OHEM)method and the loss function to improve and optimize performance of the model,and prevents overfitting with the help of the label smoothing method.Secondly,it adopts the method of backbone replacement and model pruning to realize model compression and maintain real-time inference of the algorithm.Finally in aspect of fine-grained classification,it performs network pruning and multilabel classification for SmallerVGGNet to ensure its lightweight running on low power edge terminals.Experimental results indicate this method is superior to traditional algorithms in coarse-grained detection and inference speed could meet real-time requirement for edge terminal application.Meanwhile,this method can also satisfy requirement for accurate classification of the objects according to different attributes in substations in terms of fine-grained.