A Novel Feature Set for Illumination Invariant Human Detection

Abstract

We, as human beings, are very good at extracting features that allow us to differentiate between different types of objects in our surroundings. This is because we are well aware of certain unique features that describe humans and help us separate humans from other living and nonliving objects. In order for the machine to achieve the same task with a high accuracy, we need to look for especially discriminant features that could help achieve maximum separation in human and non-human classes. This project develops an accurate human detection system that can detect human independent of variation in scale, posture and level of illumination. A novel illumination invariant feature set based on histogram of oriented gradients (HOG) and local phase quantization (LPQ) is proposed. In order to evaluate the performance of the proposed human detection system, it is trained on INRIA person dataset and tested on University of Central Florida’s (UCF) sports action dataset and Change Detection (CDW 2014) dataset. The area under the Receiver Operator Curve (ROC) of the detector turned out to be 0.781 for UCF dataset and 0.826 for CDW 2014 dataset. These results indicate that the proposed human detection system performs comparably better than state-of-the-art detectors such as HOG, Deformable Parts Model (DPM), Aggregated Channel Features (ACF) and Local Decorrelated Channel Features (LDCF).