Pedestrian attribute recognition: Upper body clothing classification

Abstract

Pedestrian Attributes Recognition has become more prevalent and important in safeguarding the community from the crimes. It is the solution to replace the old, cumbersome method of Criminal Characteristics Identification with a more advanced, efficient and accurate framework. The widespread usage of Closed-Circuit Television (CCTV) and the emergence of Big Data prepares a perfect ground for an advanced analytic tool to dissect and understand the massive collection of video footage for various purposes. Therefore, the aim is to tackle the issue of pedestrian identification using one of the attributes, upper body clothing classification. For this purpose, P-DESTRE dataset is chosen due to the multiple attributes of the pedestrians available including upper body clothing classes. A few pre-preprocessing steps are also necessary before feature extraction such as DeepLab for background removal and AlphaPose for body parts recognition. In this paper, two major elements are used in classifying upper body clothing, type of sleeves and type of collar. The type of sleeves requires the calculation of skin over arm section pixels percentage meanwhile the type of collars needs Features from Accelerated Segment Test with Non Maximal Suppression (FASTNMS). The findings show that the classification accuracy rate of both two elements achieved a over 90% which shows the effectiveness of the two methods, thus helped to establish a framework to recognize a pedestrian based on upper body clothing.