Improved 3D Object Detection Method Based on PointPillars

Autopilot perception technology is experiencing rapid development. LiDAR provides critical information for the autopilot system by perceiving objects, roads, and behaviors in the surrounding environment of the vehicle. Aiming at the problems of low recognition accuracy of target objects and prone to missed and false detections by the three-dimensional target detection algorithm in complex scenarios, a three-dimensional target detection algorithm based on the improved PointPillars is proposed. It adopts Distance-based sampling to reduce the influence of point cloud feature loss; and uses the 3D CIoU loss function to improve the accuracy of the PointPillars algorithm. Compared with the original Point-Pillars network, the average accuracy of the improved algorithm on the categories of cars, pedestrians, and cyclists has increased by 3.7%, 5.9%, and 5.7% respectively, demon-strating the effectiveness of the proposed method.