Research On Pedestrian Detection Based On SSD

At present,intelligent development has become a trend,and computer vision is an important research field of artificial intelligence,because of its very wide application and can bring great commercial value,it has become the favor of industry and academia.As a classic problem in object detection,pedestrian detection has great scientific value in many practical application scenarios,including security,unmanned,monitoring and robot.In the process of the actual scene pedestrian detection,due to the complex and changeable environment,people's wearing,the change of external light and the change of background,all kinds of body posture,occlusion and other factors,the appearance gap of pedestrians is very large.Accurate and rapid pedestrian identification and positioning is still a difficult and challenging research task.Aiming at this phenomenon,this paper studies the convolution neural network model based on SSD(Single Shot MultiBox Detector),and the improved algorithm improves the detection speed and accuracy compared with the current classical SSD algorithm.The specific work of this paper is as follows:(1)For improving the training speed of the algorithm,this paper improves the basic network part of the SSD network.in order to reduce the complex computation of convolutional neural networks,convolutional kernels with different sizes are used to limit the number of input signals,and the network dimension is reduced by adding convolution layers with single channels.Because the distribution of input data in each layer will change during training,resulting in a certain degree of information loss,batch normalization is added at the output end of each convolution layer,so that the input distribution of the next layer neural network remains the same,thus speeding up the convergence of network training,while improving the detection speed of convolution network.(2)For improving the detection accuracy of the algorithm,this paper optimizes the feature extraction method of convolutional neural network in SSD algorithm,so that the final output features can better express the feature information of each dimension of the input image.there are three basic processes of this method,which are generated from the bottom-up of different dimension features,the top-down feature supplement enhancement,and the association expression between the convolutional neural network layer features and the final output of each dimension feature.The depth of neural network,convolutional kernel size and feature layer selection have a great impact on the performance of target detection.in this paper,based on the SSD of target detection algorithm,we propose a pedestrian detection method based on improved SSD of sparse connections and multi-scale fusion.this algorithm achieves good performance both training speed and detection accuracy.Through comparing the experimental data on the PASCAL VOC and CUHK Occlusion image data sets,it shows that some of the optimized designs adopted in this paper have higher accuracy than the original algorithm,and the detection speed reaches 31 fps to meet the real-time requirements,and has the characteristics of real-time performance certain application value.