A Comparative CNN based Analysis of Classification and Detection of Potholes

Autonomous vehicle (AV) systems can comprehend driving scenes has increased due to improvements in vision-based techniques. AV systems need to identify multiple detections like potholes, pedestrian, curb, road lanes, vehicles, traffic signs, signals etc., accurately and flawlessly for the fast-growing driverless vehicle technology. The prime motive of this research work is to investigate one of the major self-driving problems called pothole detection system. Autonomous vehicles must understand and detect objects based on reliable representations of their features. There are several different reasons why potholes occur on roads, including structural aging, heavy rainfall, dense traffic, thin or weak substructures, etc. It is increasingly difficult to maintain and assess such roads on a regular basis. Current methods have a limited ability to detect potholes precisely. This research work, compares the state-of-the-art deep learning Convolution neural network-based architectures VGG19, ResNet50, and DenseNet121 networks to choose the most suitable benchmark model for classification and detection of potholes using a transfer learning technique. The proposed model architecture, the test accuracy of ResNet50 was 94.6%, the test accuracy of DenseNet121 was 95.71 %, and the test accuracy of VGG19 was 94.29%. In comparison to the other two architectures, DenseNet121 achieved the best results.