Improving Detection Performance of Helmetless Motorcyclists Using the Combination of HOG, HOP, and LDB Descriptors

Abstract

A significant number of motorcyclists that do not wear helmets lose their lives during a traffic accident, one of the major causes of death globally. This led to the design and development of a system capable of detecting helmetless motorcyclists in real-time to reduce the number of deaths. Generally, this system consists of 3 subsystems, namely moving object segmentation, motorcycle classification, and helmetless head detection. The Histograms of Oriented Gradients (HOG) descriptor has been used in preliminary studies, which resulted in fast computation time and high accuracy. However, this descriptor was less effective when applied to images with varying lighting and was unable to distinguish local pattern features. Therefore, this study proposed a new descriptor algorithm, namely Histogram of Oriented Phase and Gradient-Local Difference Binary (HOPG-LDB), which combined the HOG, Histogram of Oriented Phase (HOP), and Local Difference Binary (LDB) descriptors. The HOP was used to enhance the inability of the HOG to be effectively used in detecting images with varying lighting, and the LDB was used to extract local pattern features. The results showed that the proposed method can improve the accuracy of motorcycle classification and helmetless head detection compared to HOG, HOP, LDB, HOG-HOP, HOG-LDB, and HOP-LDB descriptors. Furthermore, the motorcycle classification accuracies of the proposed method were 97.05%, 97.25%, and 99.35% for the JSC1, JSC2, and database1 datasets. Meanwhile, the helmetless head detection results of the proposed method were 71.21%, 66.63%, and 91.73 for the JSC1, JSC2, and database2 datasets.