A Binocular Vision Vehicle Distance Estimation Method For 3D Detectio

With the development of autonomous driving technology,the vehicle is more and more demanding for ranging.Most self-driving vehicles use expensive sensors to measure distance,such as lidar.At the same time,cameras have become an attractive choice because of their relatively low price.Currently,most vision-based vehicle distance estimation methods use monocular 2D detection to obtain the license plate or body size of the target vehicle and the corresponding camera projection information for vehicle distance estimation.To improve the accuracy and robustness of the ranging results,we propose an end-to-to-end vehicle distance estimation method based on binocular 3D detection.The actual area of the vehicle and the corresponding projected area in the image are obtained by the 3D detection method,and then a geometric model of the regional distance from the camera projection principle is established to estimate the distance.The main research work of this paper is as follows:(1)This paper expounds the basic concept of image object detection and the development process of object detection network,and studies the classical algorithm and transfer learning advantages and basic ideas;Multiple proposed vehicle distance estimation methods are reviewed.(2)A vehicle classification network is optimized.Using the idea of transfer learning,EfficientNet-B0 is selected as the backbone network,and the filtered CompCars dataset is used as the training set,before finally obtaining our vehicle classification network.The EfficientNet-B0 is improved by simplifying and introducing the ECA module.In the process of building the classification framework,three strategies are combined:combing with label smoothing and unique hot coding label technology,embedding Bi templated login loss function and random discarding node strategy.The experimental results show that the three strategies can improve the network detection accuracy.(3)A vehicle distance estimation method is proposed.A model for binocular 3D vehicle distance estimation is constructed to successfully obtain the real 3D size information and projected 3D size information of the target vehicle and calculate the distance of the target vehicle.(4)The accuracy and robustness of the model are tested by experiment from distance,truncation,occlusion and perspective.