Semester

Summer

Date of Graduation

2026

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Brian Woerner

Committee Member

Andrew Nix

Committee Member

Amr El-Wakeel

Abstract

This work explores the implementation of a real-time LiDAR processing model for autonomous vehicle design. Automotive perception systems are becoming increasingly important as Advanced Driver Assistance Systems (ADAS) and Connected and Automated Vehicle (CAV) technologies continue to progress toward higher levels of autonomy. Reliable perception is paramount for autonomous features, such as adaptive cruise control (ACC), where identifying and tracking a lead vehicle is a safety critical task.

Many production ADAS perception systems primarily rely on radar and camera sensors for object detection and sensor fusion. These sensors work well to achieve classification and accurate distance detections; however, they are easily corrupted by weather conditions. Fog, rain, glare, and poor lighting conditions can significantly degrade camera-based perception, leaving sensor fusion more dependent on radar measurements. While radar provides reliable range and range-rate information, it is generally less effective than camera or LiDAR for object classification. LiDAR sensors provide high-resolution three-dimensional spatial information that can greatly improve object localization and environmental understanding. As LiDAR technology continues to become more capable, compact, and affordable, real-time LiDAR processing is becoming a more practical option for research and production-oriented automated vehicle platforms.

The work presents the development, training, and evaluation of a real-time LiDAR-based vehicle detection pipeline utilizing the PointPillars network architecture. The system was implemented within the West Virginia University EcoCAR CAV controller using Ouster LiDAR sensors, the RTMaps middleware environment, and an OpenPCDet-based inference pipeline running on a dSPACE AUTERA platform equipped with an NVIDIA A5000 GPU. A custom dataset was created with logged vehicle data using the CVAT labeling toolset to identify vehicles, pedestrians, and trucks using 3D bounding boxes. This dataset was used to train a detection model tailored to the team’s sensor configuration.

The trained model’s performance was evaluated using logged driving data to determine its ability to detect vehicles in real world point cloud scenes and to assess its suitability for real-time deployment. Results showed that the trained PointPillars model produced promising vehicle detections and improved behavior when compared to pretrained models. The pretrained model evaluated in this work was trained on data with a different LiDAR configuration and point cloud distribution than the WVU 128-channel Ouster setup. This domain mismatch likely contributed to the poor object detection behavior experienced during inference with the pretrained model.

This LiDAR processing system was integrated into the team’s RTMaps perception pipeline, demonstrating the feasibility of producing LiDAR-based object detections at a rate suitable for downstream tracking, visualization, and future sensor fusion applications. The results of this work demonstrate that PointPillars is a practical approach for real-time LiDAR processing on an automotive research platform.

Download

Share

COinS