Deep Learning based Traffic Surveillance for Precise Road Maintenance Prediction using Load Analysis

Abstract

Road deterioration not only impacts the traffic flow on large scale but also causes heavy budget cuts in transportation world wide. If the deterioration can be predicted before happening or at earliest stage then it will not only maintain traffic flow smoothly but also it can be dealt with in much less budget. When a road pavement is design the Equivalent Single Axle Load ESAL is computed, to provide an estimate for road lifetime. Depending on traffic flow a static growth per year is added. But since all the procedure is done manually so it requires a lot of time and effort. This study focuses on estimating the ESAL automatically and proposed an end to end multi stage trainable framework. It majorly consist of three parts including Vehicle Detection, Axle Counting and ESAL Estimation. In first part for vehicle counting we used a two stage detector with Faster-RCNN. For this stage we used Indian Driving Dataset IDD. For the second stage we needed to detect number of axles in trucks as the axle load for trucks varies with number of axle. The axles in a truck can vary from two to six and with those axle load can vary from 23 tons to 69 tons. For second stage we created our own dataset as there was no dataset available for detection of axles in a truck. The dataset is named as Truck Tyre Detection Dataset TTDD. We used a YOLO based Single Stage Detector to detect and count number of axle in a truck. The third phase is the final one which computes the ESAL based on vehicle type and count from last two stages. For Vehicle detection phase, we acquired overall mean IOU of 0.81 and 0.91 for truck class only. For second phase, truck tyre detection we got Mean Average Precision (MAP)50 of 0.995. For end to end testing we got accuracy of 77.11% for calculating design ESAL. In conclusion this research provides a very effective solution for advance road deterioration estimation by counting total axle load that is being passed from the road.