Three-Dimensional Object Detection in Point Clouds with Multi-Stage Proposal Refinement Network

Three-dimensional object detection in point clouds serves a vital role in autonomous driving and robotics. Point Clouds provide a vivid representation of 3D data that enables reliable object detection by acquiring the spatial distribution of points in a scene, facilitating the localization and identification of the objects within three-dimensional space. Precise localization of the objects remains challenging, particularly for moderately visible objects which attributes to inconsistent quality proposals. To tackle this, this paper presents a multi-stage proposal refinement network to generate the qualitative predictions. The research contribution is, first to improve the quality of proposals in partially visible objects, the model is integrated with 3D Resnet backbone through the refinement module at various stages. Second, to improve the quality of predictions, a confidence-weighted box voting mechanism is incorporated ensuring the precise bounding box detections. Experimentation analysis was carried out on the KITTI, NuScenes and the custom LIDAR datasets. Notably, the proposed method achieves an average precision of 82.45% for Car class, 44.94% for Pedestrian class and 66.12% for Cyclist class on the moderate category of KITTI dataset, but in the hard category with high occlusion need to be improved. On Nuscenes dataset, the model achieved mAP of 66.2%. In custom dataset, 2739 training frames, 342 frames for validation, and 343 frames for testing were taken which achieved an average precision of 82.40% for Car, 44.10% for pedestrian and 67.90% for Cyclist. The results indicate that multi-stage refinement network enhances to perform the object detection precisely, which is critical to localize and detect the target in autonomous driving and robotics.