关键词: 深度神经网络, 视频监控, 特征融合, 特征图生成策略, 人体检测

Abstract:

In the video surveillance scene,due to the influence of factors such as complex background,multi-posture and occlusion,existing human body detection algorithms have problems such as low accuracy and weak model generalization ability.In response to the above problems,we have designed a detection network feature fusion method and feature map generation strategy based on the feature image pyramid and multi-scale receptive field theory.By relying on the lightweight feature image pyramid technology and combining optimization methods such as data enhancement,anchor box matching strategy and occlusion loss function,we have further proposed a human body detection algorithm EFIPNet based on the deep neural network.Meantime,in order to fully verify the effectiveness of the EFIPNet algorithm,this paper establishes 4 diversified video surveillance scene data sets,which involves a total of 50 common human body postures.The validation of the algorithm shows that the human detection network we have designed can effectively improve the detection accuracy of the human body,and achieve accurate and real-time human body detection in complex monitoring scenarios.In addition,in order to verify the effectiveness of different modules in the EFIPNet algorithm,we have used the ablation research method to analyze the influence of the main modules in the network on the performance of the human body detection model.On the Person dataset,compared with the SSD detection algorithm,the EFIPNet algorithm improves the detection accuracy of human targets by 4.34% while maintaining the detection speed of 45 frames per second.

Key words: deep neural network, video surveillance, feature fusion, feature map generation strategy, human body detection