AUTOMATIC DETECTION OF ROAD ANOMALIES USING COMPUTER VISION AND DEEP LEARNING رسالة دكتوراة

Abstract

Automatic road anomaly detection and recognition systems are essential due to their effect on the comfort and safety of drivers and passengers. Drivers should be aware of their routes’ bad road conditions and anomalies to avoid accidents, reduce the possibility of car malfunction or damage, and take the most appropriate route to their destinations. This led to a greater interest in research in automatic detection and recognition of road anomalies. Different techniques have been developed for automatic road anomaly detection and classification. The related studies can be categorized into accelerometer-based techniques and vision-based techniques. Both methods have problems. As for the accelerometer-based methods, the car’s vibration causes errors in the accuracy of the detection process. As for the image-based methods, their problems are represented by the lack of road anomalies dataset. In addition, the failure to calibrate the camera installed on the car leads to image distortion and loss of important data. This research created a new dataset from images of Tulkarm city roads taken using a DJI Mavic Air2 drone. A total of 15,326 images were extracted to create the dataset. The dataset consists of four classes: 4781 images of cracks, 4196 images of potholes, 3475 images of manholes, and 2874 images of speed bumps. All images are in 4K resolution and free from distortion, noise, or blur. Moreover, two novel deep-learning models were developed automatically to detect and clas sify all the different types of road anomalies, such as potholes, cracks, speed bumps, and man xi holes. The first model is called Dynamic Similar R-CNN (DSR-CNN). This model employs graph segmentation, graph similarity, and dynamic programming algorithms to reduce the num ber of proposed regions and increase the detection and classification speed without affecting the accuracy. The results showed that DSR-CNN significantly reduces the number of candidate regions compared to the selective search algorithm employed in R-CNN and fast R-CNN. The developed technique proposed, on average, 8% of the segments proposed by the selective search algorithm. The DSR-CNN model achieved an average speed of 0.1173 seconds per frame with a mean average precision (MAP) of 82.82%. The second model, called Dynamic Generative R-CNN (DGR-CNN), utilizes graph seg mentation, graph similarity, dynamic programming for the region proposal phase to improve detection speed, and DCGAN technique to enhance the proposed regions and thus improve the accuracy of recognition and classification. This model achieved a mean average precision of the proposed method of 94.85% with a speed of 0.262 seconds per frame, which is considered a substantial improvement in detection and classification accuracy compared to the accuracy resulting from other state-of-the-art road anomalies detection and classification research. The increase in accuracy is achieved without significantly compromising the speed of faster R-CNN