Events Calendar

Phd Defense - Weijun Wang

PhD defense: Weijun Wang

Title: Tracking Multiple Articulating Humans from a Single Camera

Time: 2:00PM -3:30PM
Location: Powell Hall of Engineering(PHE) 631
Dissertation Committee:
Chair: Professor Ram Nevatia
Suya You
C.-C.Jay Kuo

Abstract:
-----------------------------------------------------

Monocular multi-target tracking aims at locating multiple targets, maintaining their identities across frames and estimating their motion trajectories from a single camera view, which is an important problem with many applications such as automatic surveillance and video retrieval. In particular, humans are often the most concerned targets as daily activities and events in real scenes usually involve human participants. Even though some fairly significant advances have been made on pedestrian tracking in recent years, the problem of tracking multiple humans towards higher-level reasoning is still far from solved. For example , humans might move in groups in real scenes and important social context features have not been effectively explored by the usual simplification that targets' trajectories are independent. Most importantly, unlike well-studied pedestrian detection, articulated human detection remains a challenging task which makes the existing pedestrian tracking approaches less effective on videos with multiple articulating humans. In this work, we focus on exploring important online learned appearance and social context cues to improve tracking performance on pedestrians as well as articulated humans.

As pedestrian tracking is the foundation of the proposed approach, we first propose to improve its performance by considering social context. We propose a general quadratic formulation to incorporate social dependency into a global optimization problem to improve multi-target tracking accuracy. To ensure the tracking efficiency, we show an approach to convert the new binary quadratic programming formulation to a semidefinite programming problem under convex relaxation, which can be efficiently solved by off-the-shelf methods. With the new formulation, we propose to consider a few simple common trajectory dependency factors, which can be efficiently inferred online to improve tracking performance, especially in semi-crowded scenarios. In scenarios where no trajectory dependency can be explored, our solution is the same and as efficient as those classic linear optimization formulations. Experimental results on standard datasets show the advantages of our approach over state-of-the-art. Moreover, this new formulation provides a general framework to consider various useful high order information to improve multi-target tracking.

To address the problem of tracking multiple articulating humans from a single camera, we propose a hybrid framework. Our method incorporates offline learned category-level detector with online learned instance-specific detector as a hybrid system. To deal with humans in large pose articulation, which can not be reliably detected by off-line trained detectors, we propose an online learned instance-specific patch-based detector, consisting of layered patch classifiers. With extrapolated tracklets by online learned detectors, we use the discriminative color filters learned online to compute the appearance affinity score for further global association.

Experimental evaluation on both standard pedestrian datasets and articulated human datasets shows significant improvement compared to state-of-the-art multi-human tracking methods.

Location: Charles Lee Powell Hall (PHE) - 631

Audiences: Everyone Is Invited

Contact: Lizsl De Leon