Research On Vehicle Distance Measurement Based On Monocular Camera In Multiple Scences

With the continuous increase of car ownership,the importance of traffic safety is increasing.In order to avoid huge casualties and economic losses caused by traffic accidents,various Advanced Driving Assistance Systems（ADAS）have been proposed.Front vehicle ranging is an important function in ADAS technologies such as adaptive cruise control and forward collision warning system,which can help drivers make accurate judgments on imminent danger in a timely manner and avoid accidents.At the same time,in bad weather traffic scenarios,due to the occurrence of blocked sight lines,traffic accidents frequently occur.Based on the above,this paper proposes a multi-scene front vehicle ranging algorithm based on monocular vision,which can solve the front vehicle ranging in different weather scenarios.First,this paper adopts a deep learning method based on a monocular camera for vehicle ranging.An unsupervised learning neural network architecture is designed to perform depth estimation on monocular images.The neural network takes the images of the adjacent time in the monocular video as the network input.Under the joint training of the depth estimation network and the camera pose estimation network,the depth information of the image and the camera pose change matrix between the images are estimated.Videos are used as supervised information to train the network unsupervised.In the depth estimation network,an encoder-decoder design with skip connections and multi-scale prediction is employed.In order to avoid the problem that the performance of the depth estimation network is degraded due to the presence of moving objects in the image,this paper proposes a masking operation that can effectively remove the motion occlusion problem.Consistency of scales when doing depth estimation.The relative depth is converted into absolute depth by the mapping relationship between the estimated result and the true distance.Comparing the algorithm in this paper with several mainstream algorithms,it shows that the algorithm in this paper has certain advantages.Secondly,in order to adapt to the ranging of the preceding vehicle in multiple scenarios and increase the generalization performance of the ranging algorithm,this paper designs a depth estimation network based on the generative adversarial network for rain and fog and other weather scenarios.The network is mainly composed of a generator network and a discriminator network.Input images of different weather scenes into the generator network,use the attention mechanism and image mask operation to extract features from the images,and use the encoder results in the normal weather depth estimation network as the network real value for training to ensure that the generator The network is able to generate image features consistent with normal weather.The final network structure is obtained by splicing the depth estimation network with the generative adversarial network.The image results of different algorithms are compared to verify the feasibility of the algorithm.At the same time,the algorithm is verified on the KITTI dataset.Within the range of 30 m,the relative error of distance measurement can be guaranteed to be less than 4.5%,and the average relative error of distance measurement is 2.34%.Finally,in order to verify the effectiveness of the vehicle ranging algorithm ahead,this paper builds a virtual scene simulation platform to verify the ranging algorithm for the vehicle ahead.Urban roads,changing traffic lights,autonomous vehicles and pedestrians in traffic flow are constructed in the simulation scene,and the scene is controlled through Python API to realize changes in weather such as rain and fog.A monocular RGB camera and a depth camera are built on the simulated vehicle,the data collected by the RGB camera is obtained to predict the distance,and the results of the depth camera are compared and analyzed.The ranging algorithm is verified in combination with the target detection algorithm in the simulation scene,and the ranging verification is carried out on about 1,000 vehicles.The overall average relative error of the test data is about 3%.The research content of this paper expands the application scenarios of the preceding vehicle ranging algorithm,and has certain reference value for the optimization and development of intelligent driving assistance systems.