ACCA-Net: Research on Point-Cloud Segmentation Based on the Fusion of Adaptive Convolution and Channel Attention

Accurate and robust environmental perception, along with semantic scene understanding, are fundamental to the operation of intelligent vehicles. Lidar, characterized by its immunity to ambient light, long detection range, and high stability, plays a pivotal role in autonomous driving systems. Semantic segmentation, a critical task in scene interpretation, involves assigning semantic category labels to individual points within point cloud data. This paper presents a novel approach to point cloud semantic segmentation leveraging lidar range images. Utilizing spherical projection as a strong spatial prior, the convolutional filters are activated at specific locations, resulting in significant variability in feature distributions across spatial positions. To improve segmentation efficiency, this study introduces an adaptive convolution mechanism. Furthermore, to address the challenge of misclassification of small objects, a feature extraction network is proposed, integrating adaptive convolution with channel attention mechanisms. This integration facilitates enhanced multi-dimensional information interaction, thereby improving the robustness and descriptive capacity of extracted features.