An attentionbased PointPillars+3D object detection
1. School of Information Science and Engineering, Changzhou University, Changzhou, Jiangsu 213164, China; 2. Department of Energy Management, Changzhou Vocational Institute of Textile and Garment, Changzhou, Jiangsu 213164, China
Abstract：To accurately recognize and locate the surrounding vehicles and pedestrians, an attentionbased PointPillars+ 3D target detection algorithm was proposed. The entire space was uniformly divided into pillars with a given resolution, and a pseudoimage was generated by extracting pointbased features from all pillars. Two attention modules were introduced to highlight and restrain the information in the pseudoimage. A convolution neural network was used to process the output of the attention module, and the single shot multibox detector（SSD） was used for 3D object detection. The evaluation results show that the parallel attentionbased PointPillars achieves good performance. Compared with the traditional PointPillars, the mAPm is increased from 66.19 to 69.95 in the bird′s eye view， and the car mAP is increased from 86.10 to 87.73. In the 3D mode, the mAPm is increased from 59.20 to 62.55, and the car mAP is increased from 74.99 to 76.25.
