Laser Point Cloud Target Detection Based on Improved PointPillars

Abstract

Laser point cloud object detection plays a key role in autonomous driving and intelligent transportation systems, therefore, particularly important is the development of reliable algorithms for 3D object recognition. The accuracy of PointPillars algorithm in identifying small targets is low, this paper proposes an improved laser point cloud object detection method. Three key improvements were introduced based on PointPillars: Softplus activation function, attention mechanism SOPA (pillar second order of point attention) and increased positive sample weight. First, the activation function in PointPillars is replaced by Softplus. As a smooth nonlinear activation function, the Softplus function improves the detection ability of the target by enhancing the nonlinear modeling ability. Second, The attention mechanism SOPA is introduced, which improves the detection ability of important objects by adaptively weighting the key information in the point cloud. Finally, by increasing the weight of positive samples, the model places more emphasis on the accuracy of positive sample classification, which may improve the detection ability of positive samples. The extensive experimental evaluation of the improved PointPillars algorithms are conducted on the publicly available KITTI dataset. Compared with the traditional PointPillars algorithm, our method has made significant improvements in the 3D AP index. The average detection accuracy of car, Pedestrian, and Cyclist categories has increased by 0.99, 2.03, and 2.37, respectively, compared to the original algorithm. From this, it can be seen that the improvement of the algorithm is effective.