Research On An Improved Ssd Pedestrian Detection Method

With the continuous advancement of artificial intelligence technology,target detection technology has developed rapidly,and pedestrian detection technology has also been rapidly developed as one of the important researches of target detection.Since deep learning has entered people's field of vision,research on target detection has ushered in a breakthrough,and a series of target detection algorithms using convolutional neural networks have emerged.Among them,Single Shot MultiBox Detector algorithm has high feasibility in real-time high-precision target detection field.In the actual pedestrian detection system,the detection performance is easily affected by factors such as occlusion,illumination,and complex background.For the problem of poor accuracy of pedestrian detection in traffic cameras and high miss detection rate of small target pedestrians,the original SSD algorithm is improved to obtain an SSD-T pedestrian detection algorithm that merges context information,which can effectively improve the speed and accuracy of pedestrian detection rate.The main research content of the paper includes the following aspects:(1)The traditional SSD algorithm is improved,and the basic structure and characteristics of the convolutional neural network are specifically studied.On the basis of the original SSD network,the shallow feature information and the deep semantic information are fused,so that the SSD model can make up for the expression ability of the shallow feature information.For the problem of insufficiency,three pedestrian pre-selection box scales are selected according to pedestrian size features to improve pedestrian feature extraction ability and improve detection performance.(2)Aiming at the phenomenon of overfitting in the network training process,the L2 regularization method is introduced,and the optimal regularization coefficient is selected through repeated experiments to enhance the generalization ability of the model.In the iterative training process of the network,the purpose of noise reduction is achieved to avoid the model overfitting occurs,effectively improving the detection rate(3)In order to increase the diversity of samples in the data set,crawling complex scenes through web crawlers to obtain appropriate pedestrian sample data,and then unifying them with the public data set to form a more effective pedestrian data set for model training and verification.The improved SSD pedestrian detection method in this paper achieves 93.1%accuracy and 90.6%recall rate on the INRIA data set,and 91.7%accuracy and 89.8%recall rate on the Caltech data set.The results show that it is consistent with mainstream algorithms In contrast,the SSD-T model has a higher detection rate under the premise of ensuring real time.