Improving yolo object detection performance in adverse weather conditions using metaheuristic algorithms for autonomous vehicles

One of the biggest problems in autonomous driving is that the object cannot be detected accurately and quickly. Especially in adverse weather conditions (AWCs), object detection is a major problem. Weather conditions such as fog, snowstorms, thunderstorms, sandstorms, etc. degrade the performance of object detection algorithms. Object detection algorithms are generally divided into two categories: two-stage and one-stage. The two-stage approach suggests potential object regions and object detection is performed in these regions. In the single-stage approach, object detection is performed by processing and analyzing all image pixels simultaneously. While the single-stage approach emphasizes speed in object detection, the two-stage approach emphasizes higher accuracy. However, there are cases where single-stage object detection algorithms such as YOLO and its later versions are better than two-stage algorithms in detecting objects correctly. This thesis is focused on improving object detection in AWCs for autonomous vehicles using YOLO versions 5, 7, and 9, a Deep Learning (DL) algorithm. The success of YOLO depends on the effective tuning of the hyperparameters used for the optimization and training of these three versions of YOLO. Optimization of hyperparameters is an open research topic for YOLO, as it is usually implemented manually. In this thesis, Grey Wolf Optimization (GWO), Artificial Rabbit Optimization (ARO), and Chimpanzee Leader Election Optimization (CLEO) meta-heuristic algorithms were applied separately to optimize YOLOv5, YOLOv7 and YOLOv9 hyperparameters. The study focused on the impact of optimized hyperparameters on object detection in AWCs using the DAWN and the RTTS datasets. The results show that the state-of-the-art YOLO models with GWO, ARO, and CLEO significantly improve object detection, especially in the DAWN dataset, which includes weather conditions of different challenges and road-only data. The overall performance of the YOLO models on object detection for AWCs improved by 6.146% with YOLOv7 + ARO, 6.277% with YOLOv7 + CLEO, and 6.764% with YOLOv9 + GWO, respectively.