(Georgia Institute of Technology, 2017-10-31)
Chatterjee, Anirban
The US Federal Highway Trust Fund spent USD 42.95 Billion in 2015 out of which the principal expenditure was the maintenance of existing highway infrastructure. To optimize the use of resources for infrastructure maintenance, regular road infrastructure condition surveys are required. Automated road condition surveys involve the use of survey vehicles to collect road infrastructure condition data and distress detection algorithms to automatically assess the infrastructure condition. These automated road condition surveys provide a safe and efficient alternative to manual road condition surveys. However, widespread adoption of automated road condition surveys is yet to be realized. One problem is the lack of robust, accurate and fast algorithms to detect pavement distresses. Pavement cracking is by far the most widespread and serious distress on road infrastructure. Although some crack detection algorithms have been developed to provide a high level of accuracy, their computation time makes them infeasible to implement in real-time, leading to the costly requirement of saving a large volume of road infrastructure condition data for processing at a later stage. Thus, a crack detection algorithm is required which retains accuracy in a wide range of pavement conditions without being computationally intensive. To meet this research need, this thesis presents a fast and accurate crack detection algorithm. A minimal path based approach has been used to develop the crack detection algorithm. The technical approach consists of the following major steps: 1) Image preprocessing to remove isolated noise; 2) Preliminary crack segmentation to minimize false negatives; 3) Crack object generation and connection to remove false positives; and 4) Refinement of the crack segmentation through a minimal path search based procedure. The Crack Detection Algorithm Performance Evaluation System (CDA-PES) has been used to validate the performance of the proposed algorithm. This thesis also compares the proposed algorithm with the previous state-of-the-art algorithm. The proposed algorithm outperforms all previous algorithms tested using the CDA-PES. Additionally, the proposed algorithm achieves on average a 36 times faster computation speed than the existing state-of-the-art algorithm and a 58 times faster median computation speed. With a median processing time of 0.52 seconds for 0.65 megapixel images on a single CPU thread, this algorithm makes accurate, real-time processing viable. The research presented in this thesis contributes significantly towards more widespread adoption of safer and efficient automated road condition surveys.