Research On Object Detection And Safety In Unmanned Driving

With the rapid development of society and economy,the number of cars is constantly increasing,the urban transportation system is becoming larger and larger,and the problems of traffic accidents and traffic congestion are getting worse.At present,artificial intelligence technology is developing at a high speed.Unmanned driving will be the main development direction of the future automobile industry.The development of intelligent transportation can alleviate the traffic pressure to a certain extent,and it is one of the ways to solve the above traffic problems.Vision-based object detection and traffic signs detection are two important modules of the environment perception layer of the unmanned driving system.Therefore,it is of great practical value to realize the detection and recognition of objects and traffic signs on the road by unmanned vehicles.The proposed object detection algorithm and traffic signs detection algorithm perform safety analysis to improve the safety of the unmanned driving system,which has important theoretical guiding significance.The main research contents of this dissertation are as follows:(1)In order to achieve the real-time and accuracy requirements of the object detection algorithm in the unmanned driving system,and the SSD network structure has a large amount of calculation,it is not well suited to the real-time object detection of unmanned driving system.So,in this dissertation,a lightweight SSD network structure is proposed,which reduces the calculation of the SSD network and improves the detection speed of the SSD network.Secondly,in order to solve the problem of poor performance of small object detection due to insufficient semantic information of the convolution feature map of the shallow prediction layer of the SSD network,a feature fusion method of traditional HOG features and features of lightweight SSD in Conv4-2 convolution layer is proposed,and then using the obtained fusion features for object detection,enhances the semantic information of the lightweight SSD in the convolutional layer,and improves the ability of the lightweight SSD algorithm to detect small objects.Finally,this dissertation also analyzes the safety of the proposed object detection algorithm based on the loss function.(2)In order to achieve the accuracy requirements of the traffic signs detection algorithm in the unmanned driving system,this dissertation proposes an improved Multi-class fully convolutional neural network algorithm.Firstly,the Multi-class fully convolutional neural network is used to detect the traffic signs.Then use the color detection algorithm and the computational connected domain algorithm to further filter the detection results.The algorithm improves the accuracy of the Multi-class fully convolutional neural network in detecting traffic signs.In addition,due to the large number of traffic signs and the imbalance of the number of images of each type of traffic signs in the data set,a data enhancement strategy based on the combination of image transformations such as appearance transformation,increased noise,and image synthesis is proposed.The enhancement strategy expands the traffic signs pictures in the data set,and provides sufficient data for the training of the network model.Finally,this dissertation also analyzes the safety of the proposed traffic signs detection algorithm based on the loss function.(3)The application system of the traffic signs detection algorithm proposed in this dissertation is designed.The system mainly provides the function of detecting the traffic signs videos or images and outputting the detection results,the function of downloading and querying the data set required for network model training,and the function of training the network model initially and retraining based on the existing network model.According to the actual system operation effect,the traffic sign detection algorithm proposed in this dissertation has high application value.