Research On Vehicle Early Warning System Based On Deep Learning And Monocular Ranging

In recent years,the development of the automobile industry has brought a lot of convenience to people’s lives,but the hidden danger brought by traffic can not be underestimated.Now the research on vehicle safety has become a hot direction.In order to improve the driver’s driving safety,this thesis proposes a vehicle early warning system based on deep learning and monocular ranging.The design of early warning system not only needs to make early warning strategy scientifically,but also needs to take the results of target detection and ranging as the data basis,which requires high real-time and reliability of target data.In the aspect of target detection,most of the current detection algorithms are difficult to meet the speed and accuracy requirements in the embedded terminal;in the aspect of ranging,the traditional monocular ranging is easily constrained by the road geometric conditions,and the camera calibration error and distortion will also have an important impact on the ranging results.In view of the above problems,this thesis has carried on the technical innovation in the aspect of target detection and ranging.(1)Aiming at the problem of the low detection speed and accuracy of the Tiny-YOLOV3 algorithm,this paper rebuilds the network structure from the compressed model,deepens the depth of the feature extraction module,increases the feature fusion module,and adds the newly designed SPP module;at the same time,the GT frame is rebuilt.Optimization of matching strategy with anchor box,improvement of loss function and optimization of NMS strategy.Experiments show that when the deletion threshold is 0.34,the model in this paper can obtain the optimal m AP on the road data set: 77.52%,which surpasses the Tiny-YOLOV3 model by 8 points.The running speed on the Zynq7020 platform has reached 21.3FPS,which is more than 3times faster than Tiny-YOLOV3,and the overall performance is better than other mainstream lightweight networks.(2)In view of the limited robustness and accuracy of traditional monocular ranging methods,this paper proposes a multi-task type convolutional neural network model to measure the target distance.The network uses two branches.Branch 1 is responsible for ROI target box regression,and branch 2 is responsible for distance regression.Five residual blocks and two full connection layers are used.In the second half of the network,the center point coordinates and the width and height information of ROI are input,and the information is fused with the upper convolution information.Through the direct target level range regression of the ROI area,the ranging error within 50 meters is basically less than 1.4 meters.When the ROI input size is 128×128×3,the ranging speed,accuracy and stability can meet the actual requirements.Compared with the traditional monocular ranging method,it has obvious advantages.(3)This subject carries out the overall design of the early warning system,and designs the overall architecture of the software after the software requirements analysis,and designs the software implementation process of the GRYOLO target detection module,monocular ranging module and early warning level judgment module.In the third and fourth chapters,the target detection and On the basis of deep learning ranging,the image target and the millimeter wave radar target are fused,and the early warning strategy is combined to carry out the scientific early warning system design.The Zynq7020 is selected as the system implementation platform,and the forward collision grading early warning hardware system architecture is designed.Finally,the early warning system is designed.test analysis.