Improved PointPillars for 3D Object Detection Based on Multidimensional Feature Fusion

Real-time, high-performance 3D object detection is crucial in autonomous driving, especially for applications such as cooperative driving, where accurate and timely detection of surrounding objects is essential for vehicle-to-vehicle and vehicle-to-infrastructure communication. To enhance the feature extraction capability of 3D target detection in LiDAR point cloud data and improve the detection accuracy of target objects, this paper improves the PointPillars algorithm. Aiming at the problem that the max-pooling operation will lead to the loss of fine-grained information, this paper proposes a point cloud module coding module for combinatorial pooling. We conducted experiments using the KITTI dataset, and the improved method resulted in higher detection accuracy than the baseline model. The average accuracy of 3D detection in easy, medium, and difficult scenarios improved by 7.43%, 2.02%, and 6.27% respectively. Meanwhile, the detection speed of the improved algorithm reaches 19.84 FPS, proving that the method effectively improves the detection accuracy while maintaining the real-time nature of the detection algorithm.