Multi-Target Tracking and Facial Attribute Estimation in Smart Environments

This dissertation presents a study on three different computer vision topics that have applications to smart environments. We first propose a solution to improve multi-target data association based on l1-regularized sparse basis expansions. The method aims to improve the data association process by addressing problems like occlusion and change of appearance. Experimental results show that, for the pure data association problem, our proposed approach achieves state-of-the-art results on standard benchmark datasets. Next, we extend our new data association approach with a novel technique based on a weighted version of sparse reconstruction that enforces long-term consistency in multi-target tracking. We introduce a two-phase approach that first performs local data association, and then periodically uses accumulated usage statistics in order to merge tracklets and enforce long-term, global consistency in tracks. The result is a complete, end-to-end tracking system that is able to reduce tracklet fragmentation and ID switches, and to improve the overall quality of tracking. Finally, we propose a method to jointly estimate face characteristics such as Gender, Age, Ethnicity and head pose. We develop a random forest based method based around a new splitting criterion for multi-objective estimation. Our system achieves results comparable to the state-of-the-art, and has the additional advantage of simultaneously estimating multiple facial characteristics using a single pool of image features rather than characteristic-specific ones.