Research On Multi-branch Convolution Of Pedestrian Detection

Pedestrian detection technology is a hot research content in the field of computer vision.With the progress of artificial intelligence technology and the increase of demand,pedestrian detection has become the basic research of many popular applications,such as video surveillance,automatic driving,artificial intelligence robots,etc.In the past two decades,the pedestrian detection technology in the field of research boom is higher and higher,especially in depth study and big data era,the depth of the convolution of the neural network appeared to target detection techniques progress fast,also drives the development of pedestrian detection,the general target in the deep learning convolution network model for the pedestrian detection obtained a good performance.The convolutional neural network has been committed to solve in the face of the pedestrian detection task scale changes affect the performance of detection,the depth of the traditional learning network model were adopted to divide and rule the ideas to solve the problem of pedestrian detection scale changes,creating a problem,when the extracted feature detection is difficult to adapt to the actual change of the scale of the pedestrians,And the receptive field corresponding to the extracted features is difficult to adapt to the change of scale.For this reason,another solution is proposed in this paper,which is to improve the model detection performance from the perspective of network receptive field.The main work of this paper is as follows:1.by learning and analyzing the SSD network,the original SSD network was improved and a new SSD detection model was designed.Firstly,feature extraction and reconstruction were carried out according to the original feature extraction method,and the idea of SSD multi-scale feature detection was changed to rebuild the feature pyramid structure;Secondly,the channel attention mechanism is introduced to deeply integrate the feature images reconstructed by feature extraction to enrich the high-level semantic information and detailed information of the extracted feature images.The experimental comparison on the dataset of Citypersons by multiple methods shows that the new SSD detection model designed in this paper can effectively improve pedestrian detection performance;2.a detailed control experiment was carried out on the receptive field,and the performance of the change of receptive field on the detection of targets at different scales was studied,and the conclusion was drawn that there was adaptability between receptive field and scale.According to the results of receptive field experiments,a multi-branch convolution block is designed to control the receptive field size of the network,so that the network can adapt to objects of different scales.3.Integrate multi-branch convolution blocks into the improved SSD network;The loss function is modified to solve the problem of the occlusion in pedestrian detection.In order to verify the effectiveness of the proposed method,experiments were carried out on ETH,INRIA and Caltech-USA datasets,and compared with several methods.The experimental results show that the proposed method can effectively improve pedestrian detection performance and achieve good results on the above data sets.Figure[29]Table[11]Reference[63]...