Clothing classification using corner features in pedestrian attribute recognition framework

Abstract

The recognition of pedestrian attributes has become increasingly important in ensuring community safety. It replaces the outdated and cumbersome method of identifying criminal characteristics with a more advanced, efficient, and accurate framework. With the widespread use of Closed-Circuit Television (CCTV) and the emergence of Big Data, an advanced analytic tool can now dissect and understand massive collections of video footage for multiple purposes. To identify pedestrians, this paper focuses on upper-body and lower-body clothing classification using the P-DESTRE dataset which provides multiple attributes for pedestrians. Prior to feature extraction, pre-processing steps using DeepLab for background removal and AlphaPose for body parts recognition are performed. The framework then classifies the collar, upper-body clothing, and lower-body clothing type by utilising a combination of Features from Accelerated Segment Test (FAST), FAST with Non-Maximal Suppression (FASTNMS), and Shi-Tomasi corner detectors. The findings indicate a classification rate of over 90% for all three elements, demonstrating the effectiveness of the method and establishing a framework for recognizing a pedestrian based on upper and lower body clothing.