The Research Of Multi-object Detection In Traffic Scene Based On Deep Learning

The object detection technology in the traffic scene is to aim at the objects of different images and videos in the traffic scene,such as vehicles and pedestrians,through the method of machine learning and deep learning for intelligent classification and identification.AS a research hotspot in theory and practice application in recent years,it has a very wide application prospect in driving support system and unmanned driving.However,due to the complexity of traffic scene,the objects of vehicles and pedestrians are affected by the intensity of light,the changing weather,occlusion and other uncertain factors,the object detection is still facing a huge challenge.At present,deep learning has achieved good results in object detection,in which the optimization of object detection algorithm based on deep learning is related to training samples.The object detection algorithm based on the deep learning is mainly put the training samples and the corresponding labels of the dataset into convolutional neural network to get a convolutional neural network model,and then the test samples of the dataset in convolutional neural network model to obtain the predicted results,and compare with the label of test labels ultimately and get the tested results.Through the research of these object detection techniques,this paper studies the accuracy and real-time of object detection.The main work of the paper is as follows:1.The existing SSD object detection network framework is used to train and fine-tuning for the domestic traffic scene,so as to form the object detection model under the specific traffic scene.A large number of different pictures under the different environments,weather changes from the actual shooting of the traffic video interception,and with tools,more pedestrians,vehicles and motorbike in the domestic traffic scene are marked,the SSD object detection network framework adjusts the parameters in the Changsha dataset.The experimental part gives the experimental results are affected by the number of different training iterations,the different output layers in the networks model,the different aspect ratio of default bounding box and data augmentation,and proves the better precision and efficiency of detection after fine-tuning network.2.The object detection network SSD-ResNet is proposed,and the SSD-ResNet improves the real-time and precision of object detection in traffic scene.Nowadays the deep residual network has better representation ability and lower training error and test error,and can make the calculation process less in the network training process,the based network VGG-16 of SSD network is replaced by residual network ResNet-26,and then train an adjust parameters of the modified network model SSD-ResNet in the KITTI dataset,the precision and efficiency of detection are improved,and compare the performance of vehicle and pedestrians of the detection in SSD-ResNet network.