Deep Learning of Improved Part-Aligned Features for Person Re-Identification

• Main Authors: Jui-Shan Chan, 詹瑞珊
• Other Authors: Sheng-Luen Chung
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/a6h6a7

碩士 === 國立臺灣科技大學 === 電機工程系 === 106 === Person Re-IDentification (Re-ID) is to recognize and retrieve a person who has been seen before by different cameras from possibly scenes covered by a surveillance system, commonly deployed at public and private premises. Re-ID poses as one of the most difficult computer vision problems owing to the enormous amount of identities involved in a large-scale image pool, with much similar appearance constrained by low resolution image, in a possibly occluded scene, etc. Global features geared for general object recognition and face recognition are far less adequate to re-identifying a same person across cameras. As such, more discriminative features are needed to identify people. In particular, part-based feature extraction methods that extract by learning local fine-grained features of different human body parts from detected persons have been proved effective for person Re-ID. To further improve the part-aligned spatial feature approach, this thesis proposes a deep learning framework to better characterize a person's complete information with the following threes highlights: First, better person detection and cropping by replacing a common person detector by a 2D person skeleton joints localizer (OpenPose) to facilitate the following part-alignment. Second, finer part segmentation: by containing both horizontal and vertical divided strips from the cropped person silhouette to cover both detected person’s feature and possible accessory belonging feature. Third, better learning network, by utilizing the particular advantages of Inception in combining features of different scales, ResNet in circumventing gradient dissipation for networks with deep layers, and SENet in retaining the most critical features. Our proposed solution has been trained and tested on the two most comprehensive Re-ID datasets and compared to reported state-of-the-art solutions: for the dataset of Market1501 (DukeMTMC-reID), our proposed solution both achieves competitive results with mAP of 85.96% (84.70) and CMC 1 of 94.30% (89.84), respectively.