Research On Human Pose Detection Algorithm Under Crowded Conditions

In recent years,with the vigorous development of video image processing technology and artificial intelligence theory,the demand for automatic analysis technology for video surveillance content is also growing.In campus,video surveillance system has covered in most teaching buildings and classrooms.Therefore,in order to build a smart campus,with the strong support of deep learning in video processing,researching on related image recognition and detection technology to improve school management efficiency has important meaning.In this context,this thesis mainly studies the human body pose detection technology of complex classroom surveillance scenes,including two aspects:Firstly,this thesis proposes a method based on improved Faster R-CNN for human pose detection.This method mainly studies the problem of human pose detection using regional convolutional neural network in classroom surveillance scenarios.First of all,due to the low quality of dataset,this thesis uses the robust Faster R-CNN network as the basic framework to extract high-quality candidate regions.Then,the merged ROI Pooling method is used to combine the high-level convolution features with the low-level convolution features,so that the merged features contain both high-resolution information and semantic information.Finally,the local feature preserving learning strategy makes feature distribution of same category become closer in feature space,which enables the whole network possess stronger classification ability.Experiments show that the proposed method outperforms current detection methods with higher accuracy.At the same time,this method also has some shortcomings.For example,with a large time cost,this method is not able to meet the requirements of real-time computation.In addition,the region proposal network may contain some false objects,leading false alarms in the final detection results.Secondly,to make up the shortcomings of the previous work,this thesis proposes a real-time human pose detection algorithm.The method is based on the real-time lightweight object detection network,and the basic network is further improved through a feature fusion mechanism.First of all,benefitting from the strong semantic information generated by semantic segmentation network,this thesis uses a semantic compensation network based on weak segmentation.This network combines the convolutional features of the compensation network and the shallow features of network to enhance the semantic information of shallow features.In addition,the global activation module of the attention mechanism increases the weights of key channels in convolution features,which and indirectly enhances its semantic information.Finally,the Inception Dilated-ConvNet is used to enlarge the receptive field and enhance the representation abilities of local information.Experimental results show that this method can achieve better detection accuracy and better real-time performance compared with the previous method.