DRIVABLE AREA DETECTION USING YOLOV7

Abstract

Autonomous vehicle, also referred to as self-driving vehicles, represent a paradigm shifting advancement in the realm of transportation, characterized by their ability to operate independently without any human intervention. Such vehicles require the usage of drivable area detection systems. In this dissertation, a YOLOv7 oriented framework using detection and segmentation was proposed to detect the safe-drivable area effectively and efficiently. The presence of numerous potholes on Indian roads is a serious problem because they increase the likelihood of accidents. Pothole detection is hence a must for drivable area detecting systems. Nonetheless, it can be challenging to tell the difference between the roadways and the potholes. HybridNets, being an advanced drivable area identification model, is limited to the tasks of lane-line segmentation and drivable region recognition, lacking the ability to detect potholes due to this limitation. To address this gap, the proposed approach suggests incorporating instance segmentation of road scenes using YOLOv7, which utilizes the E-ELAN layer in its backbone to facilitate the learning of more diverse and improved features. The utilization of various cardinalities and group convolutions in the E-ELAN layer promotes expansion and enhances the model's ability to acquire a wider range of features. Through extensive experimentation on our dataset, the proposed method achieves an impressive mean Average Precision (mAP) of 87%, outperforming the state-of-the-art HybridNets model which achieves an mAP of 67.8%. This demonstrates the effectiveness and superiority of the proposed approach in drivable area identification.