Improving offline tracking accuracy by merging broken tracks in traffic surveillance

Abstract

This thesis explores the problem of broken trajectories in offline multi-object tracking systems, particularly in the context of Indian traffic surveillance. In real-world urban traffic scenarios, occlusions, dense object interactions, and abrupt motion often cause detection-based tracking models to generate fragmented tracks. These discontinuities reduce the reliability of downstream traffic analysis and pose challenges for accurate vehicle movement reconstruction. The aim of this project is to study the causes of broken tracks and investigate potential strategies for merging them to restore trajectory continuity. A custom dataset was developed by extracting trajectories from traffic surveillance videos and categorizing them based on entry and exit margins using spatial criteria. To understand how current tracking methods handle fragmentation, existing machine learning-based offline tracking models were applied to the publicly available MOT17 dataset. The outputs are being analyzed to observe patterns in track discontinuities and understand the limitations of these models in handling occlusions and re-identification. Further analysis will focus on identifying gaps and formulating post-processing strategies to associate broken segments. This study provides a foundation for improving offline tracking accuracy and contributes to the development of more robust object tracking frameworks for complex traffic environments.