Ranging-Image-Based Methodologies to Enhance LiDAR Processing Efficiency
Loading...
Authors
CHEN, ZHIHUI
Issue Date
2025
Type
Dissertation
Language
en_US
Keywords
Background Filtering , Clustering , Deep Learning , LiDAR , Object Tracking , Ranging Image
Alternative Title
Abstract
Roadside-LiDAR enhanced traffic monitoring represents a promising roadside sensing solution for Intelligent Transportation Systems (ITS), extracting microscopic traffic trajectories from high-precision, high-resolution 3D point cloud streams for informed traffic decision-making and studies. The efficiency in the roadside-LiDAR data processing framework is a paramount factor to be considered. In real-time applications, reduced computational requirements enable the deployment of cost-effective hardware solutions, facilitating broader system implementation across transportation networks while maintaining operational reliability. For offline applications involving huge data volumes, enhanced processing efficiency accelerates processing outcomes and enables more comprehensive analyses within practical time constraints. Designing a data processing framework with less computational complexity not only reduces infrastructure costs but also promotes scalability and accessibility of roadside-LiDAR traffic monitoring solutions with varying resource capacities. This dissertation presents an efficient roadside-LiDAR data processing framework based on the ranging image, a spherical angular point cloud encoding method for conventional 10Hz mechanical LiDAR sensors. The research focuses on designing and developing a systematic suite of data processing components built upon the ranging image data structure to reduce the computational complexity by leveraging the fast searching of laser spatial relationships offered by ranging image encoding. Both online and offline processing frameworks were developed, demonstrating real-time operational performance during peak traffic hours with high amount of tracking targets while maintaining reliable trajectory quality. The online framework development begins at the raw data packet parsing level. Rather than parsing packets into point cloud coordinates, the proposed framework directly encodes the packets into ranging images according to mechanical LiDAR communication protocols. Critical processing components, including background filtering and clustering algorithm, were redesigned to achieve lower time complexity compared to conventional solutions while enabling parallelization for enhanced performance. The online framework achieves processing efficiency ranging from 67.29 to 91.46 ms/frame on a cost-effective hardware equipped with a 2.4GHz CPU during normal to extreme traffic conditions. The offline framework incorporates deep learning technologies to capture more complex traffic context information and designed to handle occlusion problems, a persistent challenge in dense urban traffic scenarios that affect trajectory quality. The framework maintains computational efficiency through an innovative lane occupancy embedding technique based on ranging image data structure to encode collective platoon behaviors. This approach facilitates the reconstruction of occluded traffic information using LSTM time sequential neural networks. Performance evaluations under GPU environment demonstrate an 18.75 ms/frame processing efficiency for in-lane trajectory extraction and a 95.3% occlusion reconstruction rate. Both frameworks present fast processing efficiency, far-reaching the real-time processing level. In the trajectory quality assessment, 1,478 trajectories are manually annotated for trajectory quality evaluation. The experimental results revealed that the online and offline frameworks can correctly identify 82.6% and 95.3% of the annotated real-world trajectories respectively without trajectory disconnection and ID-switching issues. This dissertation provides a foundational study demonstrating the feasibility and potential possibility of high-efficiency roadside LiDAR data processing frameworks using cost-effective computing hardware, and new opportunities to occlusion handling. The efficiency improvements and hardware accessibility demonstrated have profound implications for the widespread adoption of roadside-LiDAR traffic monitoring systems. By reducing the technical barriers to implementation, this research paves the way for larger-scale deployments of roadside LiDAR systems across extensive transportation networks. The framework's capability to handle both real-time operations and extensive data volumes with high efficiency positions it as a valuable tool for transportation agencies, urban planners, and researchers. Furthermore, the methodologies developed in this study establish a new paradigm for processing high-resolution traffic data, contributing to the advancement of intelligent transportation systems and smart city initiatives. As cities worldwide continue to prioritize data-driven traffic safety, operation, management and planning, the efficient processing framework presented in this dissertation offers a practical and scalable solution for next-generation transportation systems.