Publication: A human detection framework based on hog and cnnf features using visual and far infrared images
Date
2020-07-01
Authors
Chee, Kok Wei
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Human detection has become an essential feature in many current technological devices. For example, the ability to detect human is useful in a surveillance or safety alarm systems. However, human detection from images is a very challenging task. The main challenges include low accuracy, complex scene with cluttered background or illumination issue such as low lighting condition or overexposure problem. In this research, a framework to detect human by using both visual and far-infrared image is proposed. The proposed method used a single channel image as input and Power Law transformation is applied on the input image to reduce the illumination problem. Image gradient magnitude is used in selecting the Region of Interest (ROI) with the purpose of narrowing down the investigation area within the complex scene. The proposed method used the fusion of Histogram of Oriented Gradient (HOG) and Convolutional Neural Network Filters Features (CNNF) as classification feature. The features are concatenated in both visual and far-infrared image to further improve the detection accuracy under various lighting condition such as daylight or nighttime. Performance evaluation using the visual images from Caltech benchmark dataset showed that the proposed method achieved 60.11% miss rate at 0.1 False Positive per Image (FPPI). Besides, this research has also evaluated the fusion of visual and far-infrared images using a manually collected dataset. The test results showed that the proposed method can achieve 56.95% miss rate at 0.1 FPPI using both visual and far-infrared images and 73.75% miss rate at 0.1 FPPI when using visual image only. This proved that the fusion of both visual and far-infrared images can achieve better performance compared to use visual image alone for human detection.