AbstractHighways are critical infrastructures with a significant impact on economic and social prosperity. The UK's highways carry over 65% of domestic freight movement and 90% motorised passenger travel. For congestion-free travel and road users' safety, effective identification of potholes and maintaining the roads in good condition is crucial. Road potholes on the road cause inconvenience to commuters, delay in delivering product and services leading to a loss in the national GDP. Potholes on the surface of the road can lead to physical injuries and because of death. Highway maintenance is essential; however, it has become challenging to keep highways in good condition due to increasing traffic, insufficient budget, and lack of human resources. Effective detection of potholes and timely maintenance of roads is crucial for road users' health and safety. The current pothole detection methods require manual inspection of roads, performed with custom sensors installed on specially adapted vehicles. The procedure is time-consuming and labour extensive. The current pothole methods are inefficient and lagging to keep pace with the demand to keep roads in good condition. Few methods use Machine Learning models with sensory and imagery data separately to classify roads. However, the Machine Learning-based sensor data model fails to differentiate between road anomalies and hinges. The Machine Learning model with imagery data has a low defect rate when the road is full of water. The model fails to differentiate between real road anomalies and thin dark objects, similar to a road anomaly. Furthermore, in order to address the delay in road surface information sharing, the Internet of things(IoT) can be used. IoT is an emerging technology and has the potential to provide an efficient and cost-effective solution to road pothole detection. In this thesis, training and testing data were collected using a smartphone as well as downloaded from the Internet (google search) to imitate crowd data sourcing.
To address the issues of a sensory data-based model and imagery data-based model this thesis proposes a novel fusion model based on Convolution Neural Networks(CNN). The fusion model will take sensory and imagery data as two inputs and predict an output considering both sensory and imagery data. This study proposes a cloud-based crowd data sourcing method to collect data. In the cloud based crowd data sourcing method, the road users from across the world will be able to upload images of road anomalies on the dedicated cloud server. The data from the server will be downloaded at the backend to process and detect road anomalies. The proposed method will tag potholes with geographic allocation and send a notification to the road users who have opted for it. In this study, the images were collected using an iOS smartphone as a dashboard camera while accelerometer data was collected through a dedicated app on an iOS smartphone. The fusion model has achieved 87.20% precision, 92.70% recall and 89.9% F1-Score. The results show that utilising a fusion Convolution Neural Networks modelling approach with mixed input, image, and accelerometer data can produce better results. The proposed method is simple, cost-effective and computationally less extensive.
|Date of Award||Oct 2021|
|Sponsors||Interactive Coventry Ltd|
|Supervisor||Rahat Iqbal (Supervisor) & Vasile Palade (Supervisor)|