3D point cloud target recognition based on the Bi-LSTM and PointCNN network

Real traffic is complex and changeable scenes. How to accurately recognize targets in road scenes has always been one of the challenging problems in environmental perception. Lidar has been widely used in vehicle environment sensing because it has the ability of high precision ranging. To alleviate the influence of the point cloud disorder in the target recognition, most of the existing methods directly use a pooling strategy to the high dimensional- feature set to obtain local neighborhood features from the point cloud, which makes it difficult to learn global structure information. How to obtain semantic features of the point cloud is still a challenging problem. Aiming at the problem that the relationship between the features is not fully utilized, a point cloud recognition method based on Bi-LSTM (bidirectional long-term and short-term memory network) and convolution neural network is proposed in this paper. This method maps point clouds to high-dimensional space orderly and then learns the long-period dependencies between feature sequences. The experimental results show that the proposed method can achieve remarkable improvement in the case of high benchmark accuracy on the KITTI dataset.