An optimisation technique for the detection of safety attributes using roadside video data

Detection of roadside safety attributes plays an important role in road rating and improving road safety. In our previous research we have developed a technique for roadside attribute detection, however accuracy varies based on selected parameters so the main challenge is to develop a technique that can find optimal parameters. In this paper, we propose a parameter optimization technique that can optimize a Fully Convolutional Network (FCN) for road safety attribute detection. The technique incorporates an Evolutionary Algorithm (EA) as it can locate the global optimum in the search space and constructs better approximations to a solution than other optimization techniques. The aim is to optimize a number of parameters such as attribute image size, number of epochs, learning algorithms, number of layers, pooling types and activation functions. The proposed technique has been evaluated on road safety attributes data provided by our industry partner. The experimental results show that the proposed technique can automatically find the best parameters to achieve highest accuracy for road safety attributes. The best combination of parameters selected by the proposed technique for various safety attributes achieves a classification accuracy of 85-100%.