关键词: 自动驾驶汽车, 目标物检测算法, 深度预测, BEV特征融合, KITTI数据集

Abstract: In order to improve the detection accuracy of surrounding targets by self-driving vehicles, a target object detection algorithm (monocular-bird's eye view fusion, Mono-BEVFusion) with fusion of lidar and monocular image data on bird's eye view features is proposed. To construct camera BEV features, the algorithm builds a simple and efficient depth prediction network to predict the depth of camera features, which is supervised with depth truth values based on an explicit supervision method. When constructing the lidar BEV features, the laser point cloud is voxelized into a columnar grid and transformed into BEV features, and the BEV feature fusion network is designed to fuse the laser point cloud BEV features with the camera BEV features, and the fused features are inputted into the target detection framework to obtain the detection results of the target objects (cars, pedestrians, and cyclists). The Mono-BEVFusion algorithm is evaluated using KITTI dataset and real-vehicle road-harvesting data, and the experimental results show that the algorithm improves the mean average precision by 2.90 percentage points  compared with existing fusion algorithms, in which the single detection precision of the car class and the pedestrian class is improved by 3.38 percentage points and 4.13 percentage points, respectively. The Mono-BEVFusion algorithm has more stable detection effect on the occluded targets or the targets with a long distance, and can effectively avoid the leakage detection phenomenon of single sensor, which has better practical application value.

Key words: self-driving vehicles, target object detection algorithm, depth prediction, BEV feature fusion, KITTI dataset