Research On Laser Point Cloud Semantic Segmentation Algorithm Based On Deep Neural Network

With the rapid rise of computer vision,unmanned system attracted many researchers' wide attention in the robot field nowadays.The unmanned system senses its own state and its surrounding environment information through the on-board sensors,and then imparts semantic information to the environmental information around the system through robust semantic segmentation algorithm in order to realize system decision.As the “eye” of the unmanned system,environmental perception plays an important role in all process.In this paper,the semantic segmentation of lidar point cloud is realized by algorithm based on deep neural network.The main research contents of this paper are mainly composed of the following three parts:(1)Data processing: Due to the sparsity,disorder and non-uniform distribution characteristics of the point cloud data itself,the end-to-end convolutional neural network can't process the point cloud data directly.In order to solve this problem,the point cloud data needs to be pre-processed.In this paper,the sparsely scattered three-dimensional point cloud data is converted into a densely and uniformly distributed two-dimensional spherical map through spherical mapping.By this way,convolutional neural network can semantically segment the point cloud data effectively.In the process of data pre-processing,the spatial position,reflection intensity and angle information of the point cloud are preserved so that the loss of point cloud spatial information can be reduced.(2)The point cloud semantic segmentation neural network is designing based on the lightweight network structure,named SqueezeNet.And it's fused with ASPP module,crf module and reweighting layer aimed at the feature of the two-dimensional spherical image.By adding these modules,the feature extracted by the network can be enriched and fully combining the neighbor information and context features with the characteristics of point itself can improve network's performance.Besides,applying focal loss in the network optimization process can balance the imbalance of the positive and negative samples and speed up the network's convergence.Using the lightweight network as the backbone of the network can reduce the number of network's parameters and realize a real-time network.Also,a large number of comparative experiments and performance evaluations were also carried out to illustrate the applicability of the network.(3)Based on the qt program development framework,a point cloud semantic segmentation annotation tool is designed.In addition to the point cloud file loading and point cloud visualization,the neural network algorithm proposed by this paper is added to realize the automatic annotation of point cloud data.Also,manual tagging point cloud function is embedded in the tool to refine the false classification point.The region growth algorithm and minimum cutting algorithm are added to the annotation tool to realize the intelligence of artificial labeling and improves the efficiency of point cloud tagging.Based on the annotation proposed in this paper,the tagging efficiency of the point cloud can reach four or five frames per minute.