OVERVIEW

The objective of this Master Thesis is to develop a method for classifying and tracking dynamic objects in 3-D LiDAR scans. Understanding dynamic points in LiDAR data is critical for numerous applications such as autonomous driving, robotics, and surveillance, where distinguishing between static and dynamic elements is essential for accurate perception and decision-making. This thesis will focus on leveraging modern techniques in 3-D point cloud processing and/or deep learning to identify and track dynamic objects across consecutive LiDAR frames. The candidate will work with publicly available datasets (e.g., KITTI [2], NuScenes [3]) and/or real-world data collected using LiDAR sensors, with the final aim to enable real-time performance suitable for deployment in robotic systems.

OBJECTIVES

The objectives of the thesis are the following:

• Dynamic Point Detection: Develop algorithms to classify dynamic points in 3-D LiDAR scans by comparing consecutive frames. Use geometric, motion-based, and/or learned features to improve classification accuracy.

• Object Segmentation and Classification: Segment dynamic points into individual objects (e.g., vehicles, pedestrians). Classify segmented objects into categories using supervised learning models or pre-trained networks.

• 3-D Object Tracking: Implement a tracking framework (e.g., Kalman Filter, Multi-Object Tracking algorithms) to associate dynamic objects across frames. Fuse temporal information to improve trajectory estimation and reduce false positives.