Optimization of Fully Convolutional Network for Road Safety Attribute Detection

Abstract

Even though, deep learning techniques demonstrate an outstanding performance in various applications, success of deep learning techniques depends upon appropriately setting their parameters in achieving most accurate results. Therefore, in this paper, a novel Particle Swarm Optimization (PSO) based technique is introduced to optimize a Fully Convolutional Network (FCN) to recognize road safety attributes. The proposed technique optimizes parameters such as number of convolution layers, activation function, pooling type, attribute image size, number of iterations and learning algorithms. The proposed technique has been evaluated using a custom dataset prepared by extracting images from video data provided by our industry partner. The evaluation results reveal that the proposed technique can automatically determine the most suitable combination of parameters corresponding the highest classification accuracy for road safety attributes by exploring the solution space. Extensive tests in conjunction with the statistical experiments convinced that proposed approach is an appropriate technique for automatically optimizing FCN’s parameters. The best sets of parameters were obtained by the proposed technique for different road safety attributes with a classification accuracy of 92-100%.

Keywords

• Particle swarm optimization
• deep learning
• fully convolutional neural networks
• road safety attributes