Collection and Analysis of Datasets for Low-Light Vehicle Detection.

Many vision-based traffic monitoring systems are limited to daytime detection and have difficulties at later times of the day where there is low-illumination. Several studies have explored the issue of low-light object detection in different domains, but there is still a lack of understanding on the use of low-light data within the domain of vehicle detection. Thus, this research explored the use of vehicle data covering different lighting conditions and camera angles with regards to improving the performance of vehicle detectors during these aforementioned difficult conditions. A low-light dataset was produced consisting of common vehicle types found in the Philippines containing six different vehicle classes with a total vehicle instance count of 46,669. The effects of using different data variations for model training were explored in terms of their contribution to low-light vehicle detection performance. This research provides insight on how dataset composition affects model performance to produce a high-quality dataset for vehicle detection during low-light conditions. The study was conclusive that in order to produce an accurate dataset for low-light vehicle detection, the dataset must contain training data similar to the environments and features unique to low-light scenes.