Enhancing Vehicle Detection Accuracy Using YOLOv8n and CLAHE-Based Image Preprocessing in Foggy and Rainy Conditions

Abstract

Most traffic monitoring systems used in Indian cities share a common problem that is overlooked; the object detection models implemented in the systems are not designed specifically for use on Indian roads. Popular object detection models like YOLO [1] are trained using existing datasets such as Microsoft COCO [5], which mostly consist of traffic images taken in the West, and therefore, do not contain many of the types of vehicles that are commonly found in India, such as auto-rickshaws, erickshaws, and magic vehicles. As a result, when these models are applied directly to Indian roads, they produce very lowquality detection results, where many of the vehicles present are not detected at all, or are misclassified, or are detected with low confidence levels. To overcome this problem, a dedicated YOLOv8n model was developed through a local dataset of 808 traffic images containing 5,431 annotated vehicle instances across seven categories of vehicles that allow to increase vehicle detection accuracy on Indian roads. In addition to the issue with using an inappropriate dataset for training purposes, this research will also investigate the impact of weather on detection performance, as there are many different severe weather conditions in India, such as severe fog in winter, heavy rains during the summer monsoon season, and extreme windstorms or dust storms of differing magnitudes. To investigate the effects of weather on traffic monitoring systems, the weather conditions were artificially recreated through the use of physics-based simulation methods. Twelve separate conditions were produced by intermingling four weather types and three degrees of severity for each type of weather, resulting in situations in which five different types of image enhancement methodologies were utilized to aid in improving vehicle identification. Each of these included the following: histogram equalization [14], CLAHE [15], dark channel prior dehazing (image dehazing) [16], image sharpening, and AOD-Net [17]. The CLAHE methodology [15] emerged as the superior performer across all methodologies tested by providing significant improvements in the detection of objects. The use of CLAHE increased the mean average precision (mAP@0.5) by 53% under extremely foggy conditions, ultimately creating a mean average precision of 0.831, which is much higher than other methodologies tested. Furthermore, the fact that CLAHE methodology is simple, fast, and does not require training makes it a useful methodology for implementation within a vehicle identification system in the real world. Therefore, due to the potential beauty of the CLAHE methodology, it presents itself as a more pragmatic and effective way to improve vehicle identification systems within the traffic environment in India than any other methodology utilized in this study. Through this study, the need for specific datasets and image enhancement methodologies to improve detection rates of vehicles within Indian traffic systems, as well as the need for transfer learning in order to improve the overall performance of vehicle identification systems in the Indian traffic system is reinforced.