Improved feature extraction method based on Histogram of Oriented Gradients for pedestrian detection

In recent years, pedestrian detection for Automobile Driver Assistance System (ADAS) is a primordial task in the smart vehicle. Histogram of oriented gradients (HoG) is one of the most effective pedestrian feature extraction approaches to the study. In this paper, an optimization of pedestrian detection based on HOG method is presented and investigated to achieve an accurate human detection system. The study of different computation steps of the standard algorithm shows the possibility of improving the system performance, specifically in the build histograms step. The main idea is to customize each bin weight according to its contribution in the pedestrian extracted features. Actually, the different bins of a HoG improved vector that encodes a single cell will not have the same weight. Indeed, after the histograms computation, we will distribute an amplification factor for each bin in order to increase the weight bins that describe the relevant pedestrian features from a side. Top of that, we were interested to decrease the bins weight that affect the irrelevant features such as, other obstacle or the image background. The classification system is performed using a linear SVM classifier which is simple and easy to implement in ADAS applications. The performance studies using MATLAB simulation, proves the effectiveness of our approach.