Hybrid and Efficient Neural Network Design for LiDAR Point Cloud Data Processing

Abstract

This work presents a hybrid neural network (NN) model for efficient and accurate semantic segmentation for 3-D point cloud data. The proposed model uses a combination of spatial point cloud features and Multilayer perceptrons (MLPs), is lightweight, and has lesser trainable parameters than many existing models with similar or worse accuracy and performance. Specifically, a processing module of moderate complexity was introduced for effectively extracting and aggregating features from point clouds. The proposed model is highly scalable as it can process many point clouds because of its lightweight nature. Additionally, the model's performance outperforms more accurate models in processing time. Due to its lightweight feature, the proposed model is useful for next-generation deep learning-enabled devices with low computational power. The proposed model offers good performance while being efficient and fast, balancing accuracy and efficiency.