Publication

Abstract : The development of autonomous vehicles and intelligent transportation systems has underscored the need for high-quality LiDAR datasets that accurately represent realworld driving conditions. However, most publicly available datasets are collected from structured traffic environments in industrialised nations, rendering them insufficient for training models tailored to India's complex and unstructured road conditions. This study introduces a comprehensive methodology for LiDAR data acquisition and processing tailored to Indian conditions. Using a vehicle-mounted Velodyne VLP-32C sensor, we collected 3D point cloud data across rural, urban, and highway environments between Vallikavu and Karunagappally in Kollam, Kerala. The dataset captures unique challenges such as mixed traffic flows, irregular lane discipline, and adverse environmental conditions like dust, fog, and low lighting. Raw data recorded in. pcap format was processed through VeloView into. pcd files and meticulously annotated using CVAT, resulting in 12,000 labeled frames across five key object classes: Car, Pedestrian, Two-Wheeler, Three-wheeler, and Heavy Vehicle. Our methodology addresses fundamental limitations in existing benchmarks, including class imbalance, occlusion handling, and scene variability. Annotation quality was enhanced through Open3D visualization to validate the spatial accuracy of oriented bounding boxes. This dataset serves as a valuable resource for training and evaluating 3D object detection algorithms under complex, mixed-traffic conditions typical of India, supporting the development of robust AI-based perception systems for autonomous navigation in unstructured and dynamic road environments.