Please use this identifier to cite or link to this item:
Full metadata record
|Smart mobility, one of the most important topics in smart city studies, aims to reduce pollution, mitigate traffic congestion and enhance safety through research into Mass Transit Systems, Individual Mobility, and Intelligent Transportation Systems (ITS). ITS are advanced applications to collect, store and process data, information and knowledge aimed at planning, implementing and evaluating integrated initiatives and policies of smart mobility. Within ITS, high-resolution microscopic traffic data (HRMTD) can be obtained by infrastructure-based monitoring systems relying on various types of sensors. In the context of traffic monitoring, the acquisition of comprehensive information presents challenges in vehicle detection, tracking, reconstruction and classification. However, many existing traffic monitoring studies cover only one or two of these challenges, and the related developments are either not state-of-the-art or inapplicable to the relatively new technology of lidar (light detection and ranging) systems that are capable of acquiring accurate 3D data in real-time for future urban traffic monitoring. This research develops a 3D lidar-based traffic monitoring system that can provide comprehensive information through an end-to-end workflow, thereby determining fundamental traffic parameters including the number of vehicles, vehicle dynamics, dimensions and types. A three-step method is employed to realize vehicle detection, in which the first two steps are moving points extraction and instance clustering. The final step, vehicle and non-vehicle classification, is implemented by both a deep learning method (PointVoxel-RCNN, PV-RCNN) and a traditional machine learning approach (Random Forest, RF). Two frameworks are proposed to perform vehicle tracking. The first aims to provide more accurate vehicle speeds via a tracking refinement module. The other runs tracking and detection in parallel so that misdetections from the vehicle detection stage can be mitigated. Vehicle reconstruction is then implemented from the perspectives of both 2D and 3D without assuming any a priori knowledge. Vehicles can be fine-grained classified into different categories such as car, van, bus and truck. The developed traffic monitoring system has been practically demonstrated using data acquired from different laser scanners operating in different urban scenarios. It has been evaluated using roadside lidar data obtained from two different panoramic 3D lidar sensors, a RoboSense RS- ii LiDAR-32 and a Velodyne VLP-16, in four real-world case studies: a road section including a round corner, a straight road section near a traffic light, a road junction and a crossroad, respectively. Based on experimentation, more than 94 % of on-road vehicles are detected and tracked with a mean speed accuracy of 0.2m/s. The average range of vehicle trajectories is increased by c. 21% from the results of different scenes based on the improved framework. The continuity of the trajectories is also enhanced and the maximum effective tracking ranges of both tested laser scanners in different traffic scenes are found to exceed 110m. The dimensions of the vehicles being reconstructed are assessed with a Root Mean Square Error (RMSE) smaller than 0.24m. Vehicles are further classified into different categories with F1 score greater than 0.90. The reported accuracies demonstrate the potential of the developed system to efficiently serve finegrained urban traffic monitoring.
|China Scholarship Council (CSC) and Newcastle University
|3D urban traffic monitoring via roadside multi-beam lidar
|Appears in Collections:
|School of Engineering
Files in This Item:
|Zhang J 2023.pdf
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.