GarudaMOT: OBB-Adaptive Online Multi-Object Tracking with Altitude-Aware Kinematics and Histogram-based Re-Identification for Drone-Based Traffic Analysis

Abstract

Traffic monitoring using unmanned aerial vehicles (UAVs) at the NADIR viewpoint poses difficulties that the traditional MOT system cannot overcome. This is because the vehicles captured from this view appear as oriented bounding boxes (OBBs) at arbitrary angles. Tracking with the standard axis-aligned tracker results in improper matches. The purely kinematic-based trackers also fail due to occlusion in the absence of any reidentification mechanism. This thesis presents a lightweight, modular framework, GarudaMOT. The proposed framework aims to address the previously mentioned challenges without introducing any deep learning dependencies into the tracking framework. Three important contributions are made through the proposed framework. Firstly, the five state of-the-art online trackers- ByteTrack, BotSORT, SFSORT, OC-SORT and TiCrossNet are adapted to work in the OBB domain using an Extended Kalman filter. An OBB-sensitive IoU matching cost and deterministic angular alignment preprocessing operations are also introduced to address any symmetry problems. The other significant contribution is the introduction of the GarudaEnhancer plugin, which introduces two functionalities to the tracker. These functionalities, Altitude-Aware Kinematics (AAK) and Orientation-Aware Spatial Histograms (OASH), have been introduced to enhance the performance of online trackers without adding any extra processing burden. The third significant contribution is the new benchmark dataset, DRASHTI-OBBTrk, introduced to capture Indian road traffic using a UAV at NADIR view. This is accompanied by a purpose-built OBB annotation toolchain, which reduces the overload of the annotator by approximately 60%. The experimental results show that the GarudaMOT’s OBB native adaptation improves the association quality over the HBB baseline. The GarudaEnhancer plugin further improves the performance by reducing the identity switches by up to 96% across the evaluated trackers.