Abstract:

As research in the field of autonomous driving has attracted much attention, driving safety has become the primary consideration. Because the point cloud scene is cluttered and the background environment interferes greatly, and with the expansion of the acquisition range, the point cloud becomes more sparse, which makes the robustness of the detection algorithm weaker. To alleviate the above problems, this study proposes a 3D object detection network based on attention mechanism and context awareness. In the point cloud processing stage, a double attention mechanism based point cloud is added to generate a point weight matrix, display and mark important point data, and suppress background noise interference. In the pseudo-map feature extraction module, the FPN(Feature Pyramid Network) module is added to reuse multi-scale features, and a Context Awareness Module(CAM) is designed to capture multi-scale context semantics. Furthermore, an Attention Guide Module(AGM) is proposed based on the source features to generate a guidance weight map with clear spatial positions, so as to alleviate the spatial ambiguity caused by redundant features. The experiments in this study are carried out on the KITTI data set test. Compared with the baseline network, the Average Precision(AP) of the proposed method for pedestrians, cars and cyclists is improved by 0.59%, 0.87% and 1.42% respectively under the difficulty index. Compared with the new baseline network, the AP of the proposed method for pedestrians is improved by 3.04%, 3.53% and 3.23% under the three difficulty levels, respectively. The results show that the proposed algorithm can effectively improve the performance of 3D object detection.

Key words: point cloud, object detection, attention mechanism, FPN, multi-scale feature, context awareness, attention guide, deep learning