Graph-based object classification techniques for autonomous vehicle radar sensors

As the automotive industry continues to evolve, the importance of visual perception systems that provide situational awareness to autonomous vehicles has become crucial. While traditional deep neural networks have been successful in solving 2D Euclidean problems over the past decade, the analysis of point clouds, especially RADAR data, presents significant challenges owing to its irregular structure and intricate 3D geometry, which are not well-suited for 2D signal processing. To address this issue, we propose two novel approaches by using Graph and Transformer based classification methods for RADAR point clouds in this thesis. The novel aspect of these studies lie in the development of an object point detection pipeline utilizing Graph and/or Transformer based methods. Each of the created models utilizes a deep learning structure from start to finish, incorporating graph convolutions on both RADAR points and feature vectors. Feature vectors are generated within the GSP framework to create a contextualized representation of the sensor data. As the movement towards comprehensive deep learning methods continues to grow in popularity, we initially present the RADAR-DGCNN and RADAR-PointNGCNN methods, both of which are constructed using graph-oriented algorithms. While our initial suggested approaches predominantly utilize graph-related methods, we improve the classification outcomes by integrating adaptations into a Transformer network within the GSP framework. To validate the effectiveness of our proposed methods, we conduct experiments using publicly available nuScenes and RadarScenes point cloud datasets. Through extensive experimentation on these challenging benchmark datasets, we demonstrate that our proposed method outperforms state-of-the-art baselines studied on the RADAR point cloud in terms of performance.