Person Re-identification In Low-resolution Surveillance Systems

The person re-identification problem is known as recognizing a specific individual captured from different non-overlapping cameras through a serial of image processing and pattern recognition techniques.It is still an unsolved problem due to the limit of the existing conditions.The challenges include the low-resolution images,complex background,incomplete contours of people and the variants in location,viewpoints,lighting condition and poses.In order to overcome the challenges,we propose two approaches on the person re-identification problem.First,we come up with a new method based on dictionary learning.For feature extraction,both stripe-level and patch-level features of the images are computed.Meanwhile,we decompose the dictionary into a view-invariant dictionary and a view-specific one to overcome the limited performance resulting from the balance between the discrimination and cross-view invariant ability.We present this multi-task dictionary learning method and show the competitive performance by comparing our results with state-of-the-art methods on two publicly available datasets.The other new proposed method is based on deep learning.The method uses the body structure information to obtain a more invariant and higher-level feature representation.And we propose a novel technique to auto weigh the different part of a human body.Different from the common loss function,we apply a Siamese structure with the classification loss and the verification loss together,which explores detailed labeled information.This work improves the experimental performance conspicuously.Both two methods are original and effective,whose ideas can also be applied to other tasks flexibly.The performance above the state-of-the-art methods verifies its value and advantages.