Single and multi-object video tracking using local and deep architectures
Abstract
Moving object tracking is a fundamental computer vision task with a wide variety of real-life applications ranging from surveillance and autonomous systems to biomedical video analysis. A robust, accurate, scalable, and high-performance multi-object tracking (MOT) system of sized objects, requires novel approaches for visual appearance adaptation and generalized learning to handle challenging cases including object shape and viewpoint invariance, illumination invariance, complex object dynamics, clutter in the scene, partial or full occlusions and degraded environments. In this dissertation, our tracking system develops two pipelines and a fusion mechanism to provide precise trajectory information for detecting moving objects of interest and trajectories for object behavior and activity-based scene understanding. For single object trajectory estimation, we extended the recognition and feature fusion based single object tracking framework called Likelihood of Features Tracker (LoFT) with color attributes, scale selection, and kernelized correlation filter modules, to improve object appearance description, adaptation to scale changes, and localization of the target within the search window. For multi-object trajectories, we propose a time-efficient detect-track-and-predict system based on a novel three-step cascaded data association scheme. M2Track combines a fast hybrid spatial distance-based gating short-term data association, a robust tracklet-linking stage using discriminative and deep-learning-based object appearance models, with an explicit occlusion handling module relying not only on motion patterns but also on environmental context including the presence of potential occluders, and other foreground and background objects in the scene. Experimental results on different international challenge benchxiii marks, tasks, and datasets ranging from wide aerial motion imagery and full-motion video to biomedical microscopy videos demonstrate the robustness and efficiency of our pipelines that reach state-of-the-art performance.
Degree
Ph. D.