Anchor-Based Spatio-Temporal Attention 3-D Convolutional Networks for Dynamic 3-D Point Cloud Sequences

With the rapid development of measurement technology, LiDAR and depth cameras are widely used in perception 3D environment. Recent learning based methods for robot most focus on image or video, but deep dynamic point cloud sequences underexplored. Therefore, developing efficient accurate method compatible with these advanced instruments is pivotal to autonomous driving service robots. An Anchor-based Spatio-Temporal Attention Convolution operation (ASTA3DConv) proposed this paper process sequences. The convolution builds a regular receptive field around each by setting several virtual anchors point. features neighborhood points firstly aggregated anchor spatio-temporal attention mechanism. Then, anchor-based adopted aggregate anchors' core points. makes better use structured information within local region learns embedding from Convolutional Neural Networks (ASTA3DCNNs) built classification segmentation tasks ASTA3DConv evaluated action recognition semantic tasks. experiments ablation studies MSRAction3D Synthia datasets demonstrate superior performance effectiveness our Our achieves state-of-the-art among as input datasets.