Road Damages detection using Deep learning

Abstract

The statistics have shown that death number is more because of road accident then terrorism, Road damages are the major cause of these accidents, timely road inspection is important to carry out. However the high cost and resource utilization made this process tedious. Lower level administration lack in the resources to timely carry out road inspection specifically in developing world. Municipalities allocate the major portion of development budget for preliminary road inspection done by most experienced road inspection manager. Road damage detector is an android application which helps road maintenance managers to carry out the road inspection more easily with less resources and time consumption. Road damage detector is a deep learning based android application which detects road damages, it classify the road damages into eight categories, so that road maintenance managers can inspect the road with less resources and time consumption by just capturing the road images in real time and specifies the location and types of damage to the road maintenance team. Initial data set has 9053 images of japan road damaged after initial training it was retrained on 1000 Pakistani road images, it took 20 hours to gather local dataset. These images are captured under different illumination conditions, initially we separated images in 8 different folders according to the type of damaged after that we draw the bounding box that contain location and damage type. Moving forward we have used the state the state of art object detection model SSD-MobileNet based on the convolutional neural network to train of dataset and recorded the accuracy on both Notebook and smartphone. At the end we have demonstrated the eight different categories of damages these eight categories are the major damage categories of road damages which are commonly found in the roads. After the final training and testing on the notebook we also developed smartphone application that can detect road damages using android smartphone. There are many solutions available in the market but none of the solution is as time and resource efficient as our proposed solution. The solution using laser light and high cost sensors are commonly available in the market. Deep learning took the solution to next stage where none of other proposed solution can reach in every aspect. Deep learning is improving fast to deal with cutting edge solutions of real world problem. Deep learning algorithms are effectively detecting object and extracting the different area of interest as it is based on convolutional neural network where the shallow layer is taking some special character of interest and passing them to the deep layers for further processing. The final application is in working form with all the proposed functionalities which we described in proposal defense. However there is room for further improvement to make this product to next level by training on rare type of damages. More data augmentation can help this model to work more robustly in different lighting conditions. Moreover this product can be upgraded to detect road damages under mud and server conditions.