Point-KGCN:A Point Cloud Target Detection Framework Based on GCN with KAN

With the advancement of intelligent vehicle technology, target detection methodologies have become increasingly diversified. Traditionally, Convolutional Neural Networks (CNNs) utilising the YOLO algorithm have dominated, but these methods suffer from the problems of costly point cloud computation and poor accuracy of small target detection. These problems affect the efficiency of target detection. To address this gap, this paper presents a novel graph neural network, Point-KGNN, designed to enhance category prediction using LIDAR data and graph representation techniques. Point-KGNN, primarily designed for autonomous driving applications, excels in predicting object categories and properties. The integration of the KAN algorithm during the graph iteration process is a key innovation. KAN has changed the way information is processed throughout the entire graph structure. Additionally, the framework incorporates multiple attention mechanisms and NMS probes to further improve performance. Experiments conducted on the KITTI3D dataset demonstrate the effectiveness of the proposed method. Specifically, the results indicate improvements of 3.56% for car targets, 2.60% for cyclist targets, and 3.14% for passerby targets in medium difficulty scenarios, compared to the existing POINT-GNN model. These findings validate the efficacy of the approach introduced in this study.