YOLO-Based Multi-Object Tracking with Reduced Order Observer for Target Trajectory Estimation

Abstract

This thesis presents an innovative and efficient approach to track multiple objects in a video sequence by integrating state-of-art You Only Look Once (YOLO) object detection framework with efficient Reduced Order Observer (ROO) based target trajectory estimator. Traditionally, Kalman Filter is utilized to estimate detected objects in video frames, however, this approach is computationally demanding due to repeated estimation of large number of state variables in each iteration. To address this challenge, ROO based trajectory estimator is proposed that focuses on estimating a subset of state vector, thereby enhancing processing speed for real-time applications. A video sequence with frame rate of 30 frames per second is fed into YOLO model which outputs bounding box defined by a state vector for each detected object in a frame. State vector elements include position (x and y axis), velocity (x and y axis), aspect ratio and height of bounding box. State estimator is required to estimate states of these bounding boxes / detected objects in next frame of video sequence. ROO separates observable states from unmeasurable states and only estimates unmeasurable states. These estimates combined with Intersection Over Union (IoU) matching are used to assign bounding boxes / detected objects to tracklets ensuring efficient tracking even in the presence of occlusion and dynamic real time environments. This approach has been validated by implementing ROO framework in MATLAB R2023b and subsequently in Microsoft Visual Studio for online tracking of objects in video frame. This research contributes to the field of multi-object tracking by providing a computationally efficient trajectory estimator with potential applications in autonomous driving, surveillance and robotics.