Research And Realization Of Vehicle-pedestrain Object Detection Algorithm Based On Visual Image

With the rapid development of China’s economy,the number of vehicles in our country has risen year by year and ranks among the top on a global scale,which on the one hand improves the living standards of the Chinese people,on the other hand,it also makes the driving safety problem in the road traffic scene become a serious topic.In recent years,the development of autonomous driving technology has shown the potential to solve this problem.As the main objects involved in traffic events,vehicle-pedestrian detection is one of the most important and indispensable techniques of automatic driving technology,which can provide driving vehicles with spatial position information of other vehicles and pedestrians in the field of vision,enhance the perception ability of the autonomous vehicle.Furthermore,it provides an important guarantee for driving safety.At the same time,in the face of complex road traffic environment,the rapid and accurate detection of vehicles and pedestrians will effectively avoid potential dangers,which demands higher requirements for the real-time capability,accuracy and stability of detection algorithms.The target of this thesis is to study and propose a vehicle-pedestrian object detection algorithm based on visual images,and to deploy this model on the unmanned vehicle platform to complete the detection of vehicles and pedestrians in the road traffic scenario.The main work of this thesis is as follows:(1)A vehicle-pedestrian object detection algorithm integrating the channel attention module was proposed.Inspired by the idea of channel attention,through the repeated use of the channel attention model in the feature aggregation stage,the information exchange between different channels of the feature map is improved,so that the network can better pay attention to the objects related to the vehicle-pedestrian detection task.Finally,the model’s positioning ability is enhanced and we can observe better detection accuracy of the algorithm and better stability under high threshold standards.(2)A vehicle-pedestrian object detection algorithm using contextual spatial information to enhance the perception ability of discriminative features was proposed.By efficiently obtaining contextual spatial information,the information transmission between the pixels of the feature map is enhanced,so that more information with discriminative features is collected by the network.Therefore,this model can better detect occluded or distant objects.The performance is also better than that of the previously proposed model.(3)After aggregating the labels of KITTI,a public dataset of autonomous driving,a series of comparison experiments between proposed models in this thesis and other state-of-the-art models were carried out.Finally,our models obtained higher detection accuracy and more stable detection effect.The lightweight model was deployed on the unmanned vehicle platform,which can initially carry out a considerable vehicle-pedestrian detection experiment in the campus environment.