International Core Journal of Engineering 2020-26 | Page 88

2019 International Conference on Artificial Intelligence and Advanced Manufacturing (AIAM) Detection of road surface identifiers based on deep learning Feng Zhang Xiaoyu Wu Chaonan Gu Communication University of China Beijing, China e-mail: [email protected] Communication University of China Beijing, China e-mail: [email protected] Communication University of China Beijing, China e-mail: [email protected] marking database including the road signs at home and abroad, and use the data expansion method to increase the database and simulate the identification of the signs under the dirty damage scene. In the mean time, the best-performing different target detection models in the current popular two-stage detection algorithm and single-stage detection algorithm are introduced into the road marking detection. By analyzing the experimental results of different models and different models under different test sets, from real-time and accurate Comprehensive analysis of the rate and other aspects, giving the advantages and disadvantages of different models, and obtaining the best road sign detection model. Abstract—Road markings are an important guarantee for road infrastructure and driving safety. In recent years, the rapid development of autonomous driving technology, and the identification ofroad identifiers has become an indispensable part of this technology. The traditional target detection algorithm in computer vision has poor regional selectivity, high time complexity and poor feature robustness. The deep learning algorithm can effectively solve these problems. At present, deep learning is widely used in audio data and images, and it is feasible and meaningful to identify commonroad identifiers. Based on the self-builtroad marking database, in view of the problems existing in the currentroad marking detection research, the latest deep learning method is used to detect the six common types ofroad signs such as the right turn arrow, the straight right turn arrow and the crosswalk, and the SSD is trained. Modifying the four target detection models of focal loss SSD, FSSD and R-FCN and testing the road surface. Through data analysis and comparison, the following conclusions can be drawn: R-FCN has higher accuracy, but slower speed; The accuracy of SSD model is low whilst the detection speed is fast. The accuracy of the SSD model after modifying the focal loss is improved compared with the SSD model. The accuracy of the FSSD model is between R-FCN and SSD, and it can also maintain a faster speed. Comparing the experimental results, the R-FCN model with 18-layer depth residual network is the best for road marking detection. II. R ELATED W ORK A. Research Status ofroad Mark Detection Traffic safety has always been a hot spot for human attention. In recent years, traffic accidents have occurred frequently, and road junctions have many accidents. Therefore, the detection and identification of road identifiers is particularly important. The traditionalroad identifiers are mainly based on a variety of image preprocessing methods, and theroad marker detection algorithm is divided into two steps: marker detection and recognition. The method proposed by Foucher P et al. to identify signs such as road arrows and crosswalks is to improve image processing [11]. Yamamoto J et al. proposed an algorithm based on road marker recognition. The algorithm first obtains the marker region based on the edge information and the method of correcting by inverse perspective transformation. The marker is classified and detected by the neural network, but the detection rate is low. Especially the orange marker [20]. S.Suchitra et al. proposed an algorithm for identifyingroad identifiers based on Hough transform. Although the identifiers obtained by this algorithm have the characteristics of constant scale and rotation, the anti- interference ability based on Hough change is poor [14]. Vokhidov H et al. proposed an algorithm for identifying dirtyroad markers based on convolutional neural networks. The algorithm uses a seven-layer convolutional neural network model. The first three layers are used to extract features and the four layers are used for classification. however, the detection speed is too slow and is not suitable for practical application [1]. In order to further improve the anti-interference and real-time performance of the detection mark, some scholars have proposed the road marking detection based on the lane line since the identifier is often inseparable from the lane line, and the area of the road marking is determined according to the lane line. Meets real- time but poor robustness. In summary, it can be known that the current research onroad marking recognition mostly stays Keywords—R-FCN, SSD, Focal loss, FSSD, road marking detection, deep learning I. I NTRODUCTION Urbanization and people's living standards are improving, the number of road vehicles are increasing, and traffic accidents occur frequently. therefore, research on the detection and identification ofroad signs has become particularly important. The self-driving car launched by Google is a successful example. The traditional road detection algorithm adopts artificial design features with certain limitations and poor versatility. Most of them use image processing methods to enhance the accuracy of road marking detection. Therefore, the road marking detection algorithm has been stagnant in practice. The effect is poor in the application. Until the introduction of deep learning algorithms into computer vision, the field of target detection has ushered in a golden period of rapid development. The depth learning algorithm is flourishing due to its adaptability, universality and robustness. Some scholars have begun to adopt convolutional neural networks for marker detection to achieve better results than traditional target detection algorithms, however, there are still some defects in real-time and accuracy. The contribution of this paper is to use the self-built road 978-1-7281-4691-1/19/$31.00 ©2019 IEEE DOI 10.1109/AIAM48774.2019.00020 66