Research And Application Of Monocular Vision Road Object Detection Algorithm Based On Convolutional Neural Network

With the development of the times and the advancement of artificial intelligence technology,autonomous driving technology has gradually become a hot spot in the field of smart car research.Among them,in order to achieve full automatic driving,we must first realize intelligent driving assistance.Road object detection is an important part of intelligent driving assistance,which can effectively improve the safety and reliability of intelligent vehicles.In recent years,the application of deep learning algorithms represented by convolutional neural networks in the field of computer vision has made the performance of the visual tasks for object detection significantly better than traditional methods.Therefore,this paper carries out the research and application of convolutional neural network in monocular vision road object detection algorithm.The main research contents of this article are as follows:(1)Research on global multi-scale road object and distance detection algorithm.In order to solve the technical problem that the accuracy and real-time of road object detection are difficult to balance and the single 2D detection dimension cannot obtain sufficient information,this paper selects one-stage object detection algorithm YOLOv3 as the research basis.The paper designs a global multi-scale road object and distance detection model and the main innovations are: 1.This paper proposes to combine the perspective projection relationship of the camera system and the YOLOv3 bounding-box prediction mechanism,redefine the loss function of the network,propose a perspective projection transformation loss,and establish a distance prediction mechanism,so that the model can identify and locate road object,while estimating the distance of the road object;2.Based on the YOLOv3 network,this paper adds two feature components: the Global Context block in the feature extraction backbone network and the Atrous Spatial Pyramid Pooling block in the multi-scale prediction process,which used to improve the representation ability of the model to the global context information and multi-scale features of the image,so as to achieve the purpose of improving the accuracy of the overall detection algorithm.The algorithm designed in this paper gets m AP of 89.28% for road object detection on the KITTI dataset and the relative error of 8.51% for road object distance estimation,expanding from a single 2D detection result to a 2.5D detection result,which can achieve high-precision and faster road target detection and distance estimation in natural scenes.(2)Design and implementation of road object detection warning system.Firstly,considering that the complex and changeable driving environment in practice will inevitably lead to the problem of missing detection frame and unstable jitter in the detection system during video stream detection,this paper designs a post-processing of road object tracking algorithm based on Kalman filter to improve the robustness of the system.Secondly,in order to reduce the parameters of the model and speed up the forward reasoning speed,this paper designs a real-time global multi-scale road object and distance detection algorithm after lightweight.Finally,this paper designs a road object detection warning system based on the above algorithm,which can quickly and accurately detect the road object and distance in the video,realize the three-level warning of road object detection,improve the safety of driving,and have a good market application prospects and the significance of ensuring social security development.