The Research And Implementation Of Head Detection In Complex Scenes Based On Deep Learning

Human body detection is widely used in the field of security,and it has great market value.However,security scenes are often complex,the density of the crowd and the complexity of the scenes lead to the human body target is likely to be occluded,so the human head that has a relatively small possibility of occlusion is selected as the detection target.In order to solve the problem of performance degradation in the case of dense targets or occlusive targets,a head detection algorithm named YOLO-OFLSTM combining convolution neural network and customized recurrent neural network is proposed.Head detection algorithm YOLO-OFLSTM is improved based on the core idea of YOLO algorithm.Firstly,high level semantic features are extracted by using convolutional neural networks such as Darknet19 and MobileNet.Then,a deformed long short-term memory only with forget gate named OFLSTM is used for feature decoding to get the prediction of boundary information by regression.Finally,an improved non-maximum suppression algorithm named Union-NMS is used to remove redundant bounding boxes and obtain the final results of human head detection.For the single detection category of human head,Union-NMS algorithm takes into account both the confidence of bouding boxes and the overlap rate between these boxes,This method effectively improves the accuracy of detection.In view of the high missed rate of YOLO-OFLSTM detection model used for small-scale targets,a feature extraction network Darknet-PPM is proposed,which is based on Darknet19 that is the feature extraction network of YOLO-v2 algorithm.By using pyramid pooling module,multi-scale features are introduced to improve the detection effect of small targets.Head detection algorithm YOLO-OFLSTM is studied and implemented based on deep learning framework Caffe and its extension.The training and testing of the algorithm is based on Brainwash and Walkingstreet dataset.The occlusion of human head targets is more complex in Brainwash,and is more dense in Walkingstreet.Experiments on these two datasets show that the algorithm is effective in dense and occlusive situations.