Real-Time Object Detection and Tracking of Fast-Moving Small Objects Using RGB-D Camera and Computer Vision Techniques

Abstract

This thesis addresses the challenge of detecting and tracking fast-moving small objects using an RGB-D camera. While existing literature extensively covers object detection and tracking, there is limited research specifically focused on fast-moving small objects. Moreover, publicly available datasets are primarily centered on sports like tennis and basketball, utilizing RGB cameras with larger ball sizes and slower speeds. This thesis proposes a system design and novel method for the detection and tracking of fast-moving small objects. Real-world data was collected in a racquetball court, and existing object detection methods were fine-tuned for small object detection. An outlier detection and correction module was introduced to improve tracking performance. A physics-based tracker was developed to forecast the ball’s position, even in the presence of occlusion or missed detections. The experimental evaluation of various scenarios demonstrated that YOLOv8 offered a balance of precision and recall, meeting real-time processing requirements. The proposed system design has applications to improve robot perception on robotic platforms and explore the fusion of physics-based and deep learning methods for real-time 3D space detection and tracking of small objects. The complete source code for this thesis is accessible at https://github.com/Prithviraj97/MSThesisworks.