Study On Distance And Velocity Measurement Method Of Front Vehicle Based On Monocular Vision

The advanced driving assistance system(ADAS)uses various sensors installed on the vehicle to sense the environments around the vehicle on the road through environmental sensing technology,which are processed by the on-board computer,such that the driver can easily understand the information to ensure safe driving.With the development of deep learning,especially in severe weather conditions,higher requirements are imposed on the real-time performance and accuracy of vehicle-assisted driving on roads.In this thesis,a monocular vision-based model for front vehicle detection and distance and velocity measurement is established.By constructing lightweight convolutional neural networks(CNN),the detection of the target vehicle in front of the moving vehicle is realized.It can greatly improve the detection speed while maintaining a high detection accuracy.At the same time,the distance and velocity measurement models are established,and the detected image information is fused to generate the relative distance and velocity of the target vehicle,which provides a basis for driving decisions.The specific research contents of this thesis are as follows:Firstly,this thesis uses a single-stage multi-regression box detection(SSD)algorithm to realize the detection of the vehicle in front.In order to ensure high accuracy and high real-time detection,the lightweight convolutional neural network Mobile Net-V2 is selected as the basis of the SSD algorithm.The network uses a deep separable convolution structure to reduce the complexity of network calculations and improve the detection speed.The inverted residual structure with a linear activation function is introduced to reduce the damage to target features in the feature extraction process and extract more high-dimensional features.At the same time,based on the PASCAL VOC vehicle data set,a large number of vehicle data with occlusion and severe weather conditions are added to improve the robustness of the vehicle detection algorithm in harsh environments.Based on the Tensorflow deep learning framework,we build and train SSD Mobile Net-V1 and SSD Mobile Net-V2 vehicle detection networks,and compare their training losses and detection accuracy.Secondly,the monocular distance measurement method based on data regression is used to calibrate the monocular camera.The vehicle feature point is selected,and the world coordinate system,camera coordinate system and image coordinate system mapping model is established.The correspondence among the vehicle feature points in the three coordinate systems is revealed,and the detected pixel coordinate of the target vehicle feature point is converted into the relative distance between the two vehicles.This method does not need to separately consider the imaging error and distortion of the camera during the modeling process,which reduces the time complexity of the model and improves the real-time performance of the ranging.At the same time,a vehicle tracking algorithm based on position threshold is used to discriminate the pixel distance between two adjacent frames of vehicle feature points to achieve real-time tracking of the target vehicle and obtain the relative velocity of the target vehicle according to the time interval.Finally,an experimental environment is built and the effectiveness of the model in this thesis is verified through going onto the road.In the ranges of the relative distance of the vehicle in front of 0-30 meters and the velocity of 0-80 kilometers per hour,the experimental results show that the detection accuracy of the model in this thesis for the vehicle reaches 83.74%,the distance measurement error is basically kept within 5%,the relative deviation of velocity measurement is maintained within 10%,and that the detection speed reaches 48.30 millisecond per frame.Therefore,the model in this thesis can accurately detect the vehicles within the field of view and display the relative distance and velocity of the corresponding vehicles in a real-time manner through the on-board computer.At the same time,the research in this thesis can provide the driver with information about the surrounding vehicles,guarantee the driver's safe driving,and contribute to the technical-level realization of driving assistance systems.