A Deep Learning Approach to Pothole Detection for Autonomous Navigation on Indian Roads

Unpredictable weather and poor maintenance affecting road conditions often lead to accidents. This makes humans not frequently realize life-threatening hazards such as potholes very soon. As the road serves as an important mode of travel from one place to another, the risk of traveling should be avoided by eradicating the potholes. To avoid this and to predict the potholes that are located far away at the correct time, the early detection of the potholes is necessary. Early detection can reduce the struggles in driving and prevent the damages that are happening to the vehicles due to the accidents. Image processing plays a vital role in the detection of potholes nowadays. Various deep learning algorithms which are employed in image processing are functioning very effectively in the process of the detection of potholes at the earliest even if it is located far away. The deep learning algorithm used in this proposed study helps in the detection of potholes earlier and enables autonomous vehicles to avoid life-threatening accidents, especially in India which has roads with poor infrastructure, potholes with high density, and different road types. This proposed study uses “Adaptive transformer CNN fusion” for detecting the potholes by functioning very efficiently. Moreover, the proposed model is compared with some of the existing models such as MLP, AlexNet, and CNN to prove that the proposed worked significantly in detecting the potholes than the existing with the accuracy, precision, and recall of $98.44 \%, 0.97$ and 0.96 respectively. The proposed model has also achieved great results in the detection of potholes by classifying the affected areas in the road accurately.