Abstract
The growth in the demand for intelligent transportation systems and autonomous driving technologies enhances the development of road perception models which interpret the surroundings in real time. For this project, we design a hybrid approach to lane and vehicle detection that illustrates the classical vision techniques within the framework of deep learning, particularly using YOLOv5. The main objective is to build a modular, interpretable, and streamlined system that operates in real-time, detecting road lanes and vehicles recognized vividly with RGB videos. In the classical pipeline, lane boundary detection exploits edge tracking, region of interest masking, and the Hough transform to obtain efficient results. The combination of these methods provides not only optimal computational efficiency but also the performance of the system in low-resource environments. At the same time, a custom dataset is created for training the YOLO-based object detection model for bounding boxes of vehicles and lanes. To the raw video frames, a set of preprocessing steps comprising lens blur correction, RGB channel boosting, and edge boost are performed as part of the dataset construction process. Perspectively corrected bird’s eye view distortion helps in generating uniform bounding boxes which are converted into YOLO format after annotation. The images are annotated and these steps are added within a single automation workflow for dataset construction, model training, and real-time inference.The system is assessed using measures including: precision, recal, mean average precision (mAP) and frame processing speed. The YOLOv5 based detector has a precision of 0.91, recall of 0.88, and mAP of 0.89, with greater than 25 frames per second inference. These values reinforce the applicability of the system for automotive real-time applications. In addition, the classical lane detection pipeline separately achieves a lane detection accuracy of 0.93 and operates at 28.4 fps, confirming its utility as a primary detection algorithm and an auxiliary automatic labelling algorithm. The outputs of the hybrid model are tested against real-life recorded videos where it shows strong performance in simultaneously detecting lane markings and vehicles. The system's modular structure enables separated construction and testing of the deep learning and classical component