Vehicle Speed Estimation and Dimensionality Profiling Using 3D Lidar and RGB Camera

Abstract

Vehicle profiling is the measure of height, width, and length of vehicle. Due to increase in the traffic highways are becoming motorways. Intelligent transport system (ITS) requires the automatically profiling vehicle and its integration with online analytics systems. Speed of vehicle is measured using speed guns (1D LiDAR/RADAR) and RGB camera. Measurement of speed using speed guns require the human involvement in whole process whereas RGB camera are unable to accurately profile a vehicle. 2D LiDAR sensors are used to profile the vehicle. 2D LiDAR does a good job at profiling a vehicle but is a single lane solution and is mounted on the top of gantry. Additionally, flow of traffic needs to be controlled so that vehicle cross the gantry exactly under the sensor and speed of the vehicle is to be previously known to find the length of vehicle. This work focuses on the speed estimation (RGB and 3D point cloud data), vehicle profiling using 3D LiDAR point cloud data and synthetic data generation for these tasks using CARLA simulator. RGB speed estimation pipeline is also prepared using distance markers on the ground to compare speed estimation results using 2D and 3D data. Bird’s eye view (BEV) of 3D point cloud data is process via complex-YOLO to get oriented bounding boxes for 3D object detection. This bounding box information is used to extract width, length, and height of the vehicle. This work proposes a gantry less configuration of LiDAR being on the side of road which covers the multi lane traffic to accurately profile and estimate speed of vehicle. RGB speed estimation yields average error of 2.04 ± 1.5 km/h compared to the 0.92 ± 0.63 km/h by the point cloud data. 3D LiDAR yields overall mean errors of 0.21 ± 0.17 m, 0.05 ± 0.09 m, and 0.02 ± 0.01 m for the length, width, and height estimation