Graph attention network for 3D object detection in autonomous driving

Abstract

Since 2007, the DARPA (Defense Advanced Research Projects Agency) Grand Challenge in Autonomous Driving has significantly surged the popularity of developing Autonomous vehicles (AVs) within the autonomous industry. Safely driving in complex and dynamic environments requires AVs to have accurate and precise localization of surrounding objects. The advancement of sensor technology, particularly LiDAR (Light Detection and Ranging) brings higher accuracy and avoids limitations associated with digital camera images. Recent Advancements in the Deep Learning (DL) models have shown good performance in LiDAR point cloud segmentation, classification, and object detection tasks. However, LiDAR data generates unstructured data as point clouds around 105 3D points per 360o sweep and it's bringing major computation challenges for modern detectors to process this large amount of data in real-time. Most existing approaches use point-based, voxel-based, and view-based methods. However, point clouds are also unstructured and sparse around objects not like image pixels. Transforming to another representation causes to loss of important details related to objects. This thesis presents our Graph Attention Network (GAT) based approach to 3D Object detection using point clouds in autonomous Driving applications. The main input to our model is raw LiDAR point clouds which contain x, y, z coordinates and intensity values. The major output from the developed model is bounding box of detected cars within a given frame. Our approach can be divided into two stages. First, we remove ground points from initial point clouds in order reduce number of points involved in proceeding steps to achieve real-time processing. The second stage, we utilize voxel downsampling further reduce points in 3D space and generate a graph using the nearest neighbor technique to apply the GAT DL model. Our model evaluates using the widely used KITTI benchmark dataset. The results indicate that our model achieves performance levels comparable to those of state-of-the-art LiDAR-based 3D detection methods