Research And Application Of Vehicle And Pedestrian Detection Algorithm Based On Convolutional Neural Network

With the rapid economic development,the number of personal motor vehicles has increased sharply,and the incidence of traffic accidents has increased.Real-time and accurate detection of road vehicles and pedestrians can help to discover potential safety hazards in advance and provide early warning to drivers,which can effectively reduce the incidence of traffic accidents.The target detection technology based on machine vision is still the main method to realize vehicle and pedestrian detection.In recent years,with the rapid development of deep learning^[1],through deep convolutional neural networks and big data support,machine vision and deep learning are combined.A large number of deep learning-based methods are used for target detection,which improves the efficiency and accuracy of detection rate.In this paper,based on deep convolutional neural network to study the vehicle and pedestrian target detection algorithm,the main research content is as follows:?1?Data augmentation method based on road vehicles and pedestrian targets.Aiming at the problems that the existing multi-category open source data sets are under-represented in specific targets and complex scenes,and the sample is difficult to cover,etc.,road images under different lighting conditions are obtained through web crawling and vehicle camera collection.An image preprocessing algorithm suitable for illumination diversity scenes is proposed,a road vehicle and pedestrian data set is established,and the experimental data is augmented.?2?Road area extraction method under complex background.Aiming at the problem of difficult target feature extraction caused by complex background in actual traffic scenarios,a road area extraction method based on unsupervised learning and lane line detection technology is proposed to narrow the retrieval range of deep convolutional neural networks and improve the detection speed of the model.?3?Design of vehicle and pedestrian detection model based on improved TINY-YOLO network.For the current deep learning-based vehicle and pedestrian detection algorithm,in the complex road scene,the accuracy rate is low,the model parameter is large,and it is difficult to deploy to the embedded system for application.A vehicle and pedestrian detection method based on improved TINY-YOLO network is proposed.The improvement method mainly includes two aspects:the improvement of the TINY-YOLO network structure to achieve the fusion and multiplexing of multi-layer features;By adopting K-means cluster analysis,multi-scale training,data enhancement,alternating training and other strategies,the training process of the model is optimized.Through comparative experiments,the corresponding test set is set to test the model to verify the effectiveness and advancement of the improved model designed in this paper for real-time detection of vehicle pedestrians in complex road scenes.The experimental results show that the improved model proposed in this paper improves the m AP of vehicle and pedestrian detection in complex scenes by 1.3%.While the detection speed meets the real-time requirements of 75 frames per second,the precision and recall rate reach respectively 94.27%and 91.75%.The test model has better robustness to different target scales,different occlusion degrees and different lighting scenes.Finally,the detection model trained under the high-performance host is implanted into the embedded system,and a better detection effect is achieved in the application of the mobile terminal.