Image-based on-road vehicle detection using cost-effective Histograms of Oriented Gradients

• Image-based vehicle detection using Histograms of Oriented Gradients (HOG) is explored.
• HOGs are found to be excellent descriptors but too costly for real-time operation.
• Cost-effective alternatives to HOG are proposed and evaluated for vehicle verification.
• Proposed descriptors achieve a better trade-off between cost and performance than HOG.
• The image classification time is reduced in a factor of 27, from 8.04 to 0.29 ms.

Image-based vehicle detection has received increasing attention in recent years in the framework of advanced driver assistance systems. However, the variability of vehicles in size, color, shape, etc. poses an enormous challenge, especially for the vehicle verification task. Histograms of Oriented Gradients (HOGs) have successfully been applied to image-based verification of objects. However, these descriptors are computationally demanding and are not affordable for real-time on-road vehicle detection. In this paper, less-demanding HOG descriptors are proposed and evaluated that significantly lighten the computation by exploiting the a priori known vehicle appearance. The proposed descriptors are evaluated on a large, public database and the experiments disclose that the computation times are reduced in a factor of more than 5, thus rendering HOG-based real-time vehicle detection affordable, while achieving detection rates of over 96%.