Vision-Based Lane and Vehicle Perception Using Deep Learning and Neuro-Fuzzy Inference

Abstract

Perception of lanes and vehicles is crucial for intelligent driving assistance as well as for autonomous vehicles' navigation. However, there are many limitations of current vision-based perception approaches which can decrease their effectiveness under adverse driving conditions such as low light conditions, occlusions, and complex traffic environments. LVP-YOLOv8-seg-ANFIS, which is a vision-based approach to lane and vehicle perception with the application of YOLOv8-seg alongside an Adaptive Neuro-Fuzzy Inference System (ANFIS) is proposed to increase perception accuracy and decision-making capabilities. Firstly, images from BDD100K dataset are pre-processed with the use of the One-Sided Box Filter (OSBF) to eliminate noise and normalize the images. Then, YOLOv8-seg detects and segments both lanes and vehicles from the images on pixel level. Perceived data is analysed further with ANFIS which provides adaptive decision-making under variable driving conditions. Efficiency of LVP-YOLOv8-seg-ANFIS is assessed with the use of Accuracy, Precision, Recall, F1-score, mean Average Precision (mAP), Pixel Accuracy, Detection Rate, Intersection over Union (IoU), and Dice Coefficient. Experimental results has shown that LVP-YOLOv8-seg-ANFIS has achieved 99.5% accuracy, 98.4% precision, and 98.43% recall, exceeding MADPD-MSFWF-YOLOP, AAUDSN-AVV-MTP , and EDA-VPMAD-CNN. Combining the perception using deep learning methods and neuro-fuzzy inference provides robust and reliable perception of the lanes and vehicles, making the system useful for intelligent driving assistance and autonomous vehicle applications.