Comparative analysis of deep learning models for real-time detection of anomalies on road pavements

This thesis aims to accurately detect anomalies such as potholes, drainage covers, and speed bumps on roads. In line with this goal, a system called Dynamic Cropping System (DCS) was developed, and its object detection-based version (DCS-OD) and Instance segmentation-based version (DCS-IS) were developed. These systems can be summarized as an anomaly detection system pipeline using two YOLO models. In DCS-OD, the first YOLO model detects the road while the second YOLO model is run on the detected road area to detect the anomaly. In DCS-IS, the road is segmented with the instance segmentation model, the segmented area is entirely isolated, and anomaly detection is performed within the segmented area. At the same time, the autoencoder was placed in the anomaly detection part of the DCS-IS to experiment with the potential of the autoencoder for real-time anomaly detection. To the best of the author's knowledge, this method has not been tried before on-road anomalies. During model training, datasets were prepared by combining 36 different sources created by the author. Augmentation techniques were applied to these datasets to address class imbalance and enhance the generalization capabilities of the models. The comparison phase began after the models were trained and integrated into the designed architectures. This study compared not using a DCS (NO-DCS), DCS-OD, DCS-IS and Autoencoder systems; the most successful system was determined as DCS-OD. In the detection of road region, this system has shown superior success by reaching 0.996 precision, 0.992 recall, 0.995 map@50, 0.958 mAP@0.50:0.95, and 0.7 ms total processing time values by using the YOLOv10n model for road detection. In the detection of road anomalies, DCS-OD has provided the most balanced performance in terms of speed and accuracy by using YOLOv10s. It has shown the following success in metrics: achieving a precision of 0.868, recall of 0.764, mAP@0.50 of 0.833, mAP@0.50:0.95 of 0.564, and an inference speed of 1.5 ms. This system stands out as the most optimum option for real-time applications because it spends a total of 2.2 ms on road and anomaly detection and shows superior anomaly detection metrics. Two prototypes have been developed by integrating this system into the mobile application and the web application, where this application works synchronously, and research has been done on the common infrastructure of these applications. While the mobile application was developed with Flutter, the web application was developed with Flask framework in Python. While the mobile application provides security measures to end users by sound alerts about detected anomalies, the web application aims to determine anomaly locations in advance using Google Street View data and send them to the mobile application. The web application can also creates regional detailed reports for users. When all these outputs made in the study are evaluated as a whole, innovative systems such as DCS-OD, DCS-IS, and autoencoder, which detect road anomalies, have been developed, compared, and tested in applications. This thesis study, which sheds light on many issues, has emerged from the effect of the amount of data in the images on the detected object to the possibility of running two computer vision models on mobile phones.