| Abstrakt: |
Road management requires huge manpower and still, there is no guarantee that how many days the road will last without any cracks. Even a small crack on road can lead to a big accident. But along with these cracks, the major problem comes with the pavements which are commonly known as potholes. Due to these deadly potholes lots of people have to lose their lives without their mistakes. We also know that India has one of the largest road transportation in the world. Even in terms of population India has the second largest population in the world. Road accidents is a major topic of concern not only for the government but also for the society. Potholes have a major contribution in the road accident. The government has taken initiative by forming several agencies and through different programs which majorly focus on developing a solution for detecting these potholes which can easily detect these potholes and send the data to road management authority. Although several algorithms are used and even several techniques are used, some of them are object detection technique. But if we are using this method in real time than it’s not easy to get clear data using low spec hardware. Even the YOLO algorithms require very high spec hardware to perform proper function. But we came up with a solution by using YOLOv4 tiny algorithm which was made by combination of SPP and FPN with CSPdarknet53 tiny [1]. Thousands of raw data were obtained in the form of datasets with the help of data augmentation, namely gamma regulation, scaling, and horizontal flip to make up for the data lacks, which were further divided for training, testing, and validation. We have compared several YOLO algorithms with includes YOLOv3, YOLOv4, YOLOv3 tiny, YOLOv4 tiny, and YOLOv4 tiny –SPFPN [2]. In conclusion we get YOLOv4 tiny-SPFPN shows approx. 3-5% better performance than others in the mean average precision and also we developed real time algorithm which detect potholes in real time. [ABSTRACT FROM AUTHOR] |
