3D Vehicle Detection Based On Deep Learning Method

With the development of artificial intelligence technology,self-driving technology has also made great progress.Self-driving cars need to properly sense their surroundings to function properly.A deep learning sensing system is used to transform the surrounding environment data into semantic information that can be applied to autonomous driving.3D vehicle detection technology is a basic part of the sensing system and can sense the size and location of surrounding vehicles.Different from traditional 2D object detection,3D vehicle detection integrates tasks such as monocular depth estimation and viewpoint estimation.The current 3D vehicle detection algorithms are divided into two types according to the types of sensors.Point cloud-based algorithms require expensive depth acquisition equipment,while monocular detection algorithms have poor detection performance.Aiming at the shortcomings and deficiencies of the existing 3D vehicle detection research,this paper studies the vehicle-side 3D vehicle detection algorithm based on a monocular camera.The entire 3D vehicle detection network consists of two parts,one predicts the 2D bounding box of the vehicle,and the other predict the 3D information of the vehicle,including position,size,and yaw angle of the vehicle.The main work of this article is as follows:1.The current 3D vehicle detection algorithms use the results of 2D vehicle detection as a priori knowledge.The existing 2D vehicle detection algorithms fail to fully capture the spatial characteristics of the vehicle.In order to improve the ability of convolutional neural network to extract features of monocular image,this paper proposes an efficient 2D vehicle detection architecture.Through transfer learning,more effective features are input to the detection model,thereby improving the performance of the 2D vehicle detection module.The 2D vehicle detection data set achieved an accuracy rate of 90.49%AP(Average Precision).2.For the lack of depth information in monocular RGB images,this paper proposes an instance depth estimation based on dense atrous spatial pyramid pooling module.Our method introduces a pyramid pooling structure and adopts a multi-scale learning strategy.By densely connecting the atrous convolution layer,the receptive field of the depth estimation module is further improved without sacrificing spatial resolution,and the spatial context features of the monocular image are effectively obtained,which can predict the depth information of each instance of the vehicle.Our method achieved a competitive performance in the KITTI 3D vehicle detection data set with an accuracy of 35.56%AP3D.3.Aiming at the inaccuracy of the regression corners in the corner estimation,a corner constraint loss function is proposed,which constrains the vehicle’s 3D bounding box from two aspects of vehicle size and yaw angle.We also explore the optimal training method and analyze the effectiveness of the improvement.Experiments show that our improvement can further improve the performance of 3D vehicle detection.