HSG-HIK

"HSG" is a light-weight integer-only feature based on Histogram of Oriented Gradients (HOG). This feature classified using LUT-based Histogram of Intersection Kernel (HIK) SVM was shown to work better than HOG-Linear SVM combination for pedestrian detection despite being computationally more efficient. HSG-HIK framework plus its hardware implementation on FPGA has been described in the following paper:

M Bilal, A Khan, MUK Khan, CM Kyung, "A Low Complexity Pedestrian Detection Framework for Smart Video Surveillance Systems", IEEE Transactions on Circuits and Systems for Video Technology, vol. 27, no. 10, pp. 2260-2273, Oct. 2017.

HSG-HIK framework was later enhanced to use LUV color space in addition to gradient orientation histograms. Moreover, an aggressive approach to select hard negative examples for SVM training was proposed. Resultantly, HSG-HIK-Plus performs better than even ACF on at least one dataset i.e. ETH. See the performance curve (MR vs. FPPI) below.

M. Bilal, "Algorithmic Optimization of Histogram Intersection Kernel SVM-based Pedestrian Detection using Low Complexity Features", IET Computer Vision, vol. 11, no. 5, pp. 350-357, 8 2017.

HSG-HIK-Lite is the latest version of this pedestrian detector. This version uses SVM cascading and multiple detectors of different window sizes to significantly speed up the process without compromising detection accuracy noticeably.

M. Bilal, M. S. Hanif, "High Performance Real-Time Pedestrian Detection using Light Weight Features and Fast Cascaded Kernel SVM Classification", Springer Journal of Signal Processing Systems, 2018.

See the performance of HSG-HIK-Lite on ETH pedestrian dataset below. Detector was trained using INRIA.

DOWNLOAD SOURCE CODE