Point cloud-based data augmentation has made significant contributions to enriching training data diversity and improving model detection performance. Previous researchers have designed some data augmentation techniques tailored to autonomous driving scenarios; however, these methods did not consider the perception characteristics of point clouds, such as distance and observation angle, in the autonomous driving environment. This limited their ability to enhance the efficiency of vehicle detection models. To address this, we propose a method called Context-aware Augmentation Network for point cloud based vehicle detection (CAA-Net). This method performs data augmentation aligned with the perceptual characteristics of point clouds and improves the performance of vehicle detection models. It utilizes Context-aware Loss (CAL) to enhance the estimation of distance-aware and observation angle-aware features, accelerating the convergence of training.When applied to Sparsely Embedded Convolutional Detection (SECOND) on the KITTI dataset, our method improved the average precision (AP) by 0.47% for easy-level vehicle detection, by 1.78% for moderate-level detection, and by 8.08% for hard-level detection. The experimental results demonstrate that Context-aware Augmentation significantly enhances the performance of SECOND.
Dong et al. (Sat,) studied this question.