Investigation of YOLO models in the detection and classification of multiple negative road anomalies

Negative-road anomalies like potholes, humps, cracks present on the road surface impart a serious discomfort to drivers and passengers. Moreover, in such a road condition if a vehicle or pedestrian suddenly cross the road it becomes a significant challenge for the driver to drive safely. Early and accurate detection of multiple negative anomalies present on the road surface is crucial to prevent road accidents and for the timely repair of roads. Conventional approaches are time consuming and lack in detecting multiple anomalies. Owing to the advantage of detection of multiple objects in an image or video, computer vision (CV) based object detection algorithms are best suited for timely detection of negative anomalies. Herein we present a comparative analysis of nine different versions of the state-of-the-art object detection models (ODMs): you only look once (YOLO) in detecting multiple negative anomalies. To evaluate the performance of the ODMs, metrics such as precision, recall, F1 score and mean average precision (mAP) are measured. The frameworks are experimented using 3391 road images including potholes, speed breakers, vehicles, and humans on different road surface conditions. Upon studying the performance of all the nine variants of ODMs i.e., YOLOv5s, YOLOv5m, YOLOv5l, YOLOv5x, YOLOv6s, YOLOv6m, YOLOv6l, YOLOv7 and YOLOv7x, we observed that YOLOv6s is outperforming others resulting in mAP: 71.2 %, precision: 74.9 %, recall: 66 %, and F1 score: 70.2 % for a reasonably small image data set containing multiple negative road anomalies. YOLOv6s took 2 hours 44 minutes for training the dataset and tested images in 0.9 ms.