Research On Vehicle Distance Measurement Model And Algorithm Based On Monocular Vision Object Detection

Real-time and accurate target distance measurement capability in the driving environment is a key technology for autonomous driving.Various sensors can be used for distance measurement.Because of low cost and rich information,visual sensors can be used as an alternative to high-cost technologies such as lidar to complete the distance measurement tasks.However,existing vision solutions require a large amount of data and computing resources,and lack stability and landing capabilities.Although some methods have high accuracy,they have higher equipment requirements and complexity.Therefore,this paper takes the deep learning neural network algorithm and traditional machine vision distance measurement technology as the research object,and proposes a two-stage distance measurement model combining deep learning and traditional vision technology,which collects visual information in the real driving environment,predicts the distance of the vehicle ahead and improves accuracy and stability of distance estimates.The main work content of this paper is as follows:(1)Constructed the Dual-YOLO neural network,which can detect the body and the front and rear of the car together.Dual-YOLO uses Kalman filter and ROI algorithm to process the detection target,effectively reducing the jitter of the ranging result and improving the accuracy of vehicle target position prediction while ensuring the speed.(2)A dynamically corrected camera geometric distance measurement algorithm is proposed,and a two-stage distance measurement model TSDistance is formed with the neural network.Based on the prior knowledge of the vehicle width,the influence of the camera pitch angle on the distance measurement is paid attention to,and a single point measurement is carried out.Calculate the difference between single point distance and width based distance,compensate the change of camera attitude angle,reduce the distance measurement error caused by vehicle bumps,and improve the robustness of the distance measurement model.(3)Collect visual and radar information in the driving environment,organize and split the data,use the TSDistance model proposed in this paper to measure the distance of the vehicle target and analyze its performance and error,and verify the model’s accuracy and operating efficiency in real driving scenarios.The distance measurement model and internal algorithm proposed in this paper combine the advantages of deep learning and traditional vision,and solve the problem of unstable distance measurement results by optimizing the target detection algorithm and dynamically correcting external parameters,thus provides a feasible idea for the future research.