A Novel Tail Light Pairing System for Detecting Vehicles at Night using Histogram of Oriented Gradient

Abstract

here are many features that can be used to detect vehicles at night. However, not all of them are relevant or useful in determining between a vehicle and a non-vehicles object. This paper summarizes some of the common but important geometry features such as vehicle symmetry check, width ratio, and area differences are used to verify vehicles. In addition to previously stated features, this paper proposed the use of histogram of the oriented gradient (HOG) as a feature in making the classifier. Edge information is not normally used as features for night scenes due to the lacking of visible edge lines of objects at night. However, through observations, it is found that close range vehicles provides enough edge information to be perceived as one and vehicles that are far away too provides a special edge characteristic that enables it be classified as a vehicle. Therefore the usage of edge information can help increase vehicle validation rate and reduces false positives. Before the validation process can be made, an initial candidate tail lights pairing process need to be first made by searching for candidate tail light pair that satisfies certain general constraints. In other words a hypothesis is first made by selecting every possible pair that match a simple rule such as checking whether the pair of lights are located under the same horizontal line before further validation test is made to draw a more conclusive decision. Random forest is chosen as a tool in constructing the classifier for its speed, robustness, and simplicity. This paper assumes that the candidate tail lights are provided as inputs and the classifier proposed here functions as a pairing device to accurately pair detected tail lights of a same vehicle. Candidate tail lights can be extracted by using algorithms as simple as red color extraction. However, a better candidate tail light extractor will reduce the initial amount of regions to be validated later and will thus improve performance.