Weight analysis for various prohibitory sign detection and recognition using deep learning

  • PDF / 2,798,343 Bytes
  • 19 Pages / 439.37 x 666.142 pts Page_size
  • 25 Downloads / 280 Views

DOWNLOAD

REPORT


Weight analysis for various prohibitory sign detection and recognition using deep learning Christine Dewi 1,2

& Rung-Ching Chen

1

& Hui Yu

3

Received: 16 March 2020 / Revised: 20 July 2020 / Accepted: 31 July 2020 # Springer Science+Business Media, LLC, part of Springer Nature 2020

Abstract

Traffic sign recognition is meaningful for real-world applications such as self-sufficient driving, traffic surveillance, and driver safety. However, traffic sign recognition is a hard problem because different sizes, illuminations, and noises affect the sign detection and recognition. This work recognizes Taiwan’s prohibitory signs using deep learning methods. First, we develop a traffic sign database since there is no such kind of database available in Taiwan. Next, we adopt three different You Only Look Once (Yolo) networks (Yolo A, Yolo B, and Yolo C) and three various Yolo V3 SPP networks (Yolo D, Yolo E, and Yolo F) for prohibitory sign recognition. Finally, we conduct the comparative experiment of Yolo V3 and Yolo V3 SPP with different weights provided by the darknet framework (the best weight, the final weight, and the last weight). Experimental results show that the mean average precision (mAP) observation of all models that the Yolo V3 SPP is better than other models. Yolo D took the optimum average accuracy at 99.0%, followed by Yolo E and Yolo F 98.9%. The accuracy of Yolo V3 SPP is growing within the detection time, but it needs more time to identify the sign. Keywords Weight analysis . Object detection . Traffic sign recognition . Spatial pyramid pooling . Yolo V3

1 Introduction Computer vision-based traffic sign detection and recognition have been studied for some purposes, such as Advanced Driver Assistance Systems (ADAS) [34, 35], Auto Driving * Rung-Ching Chen [email protected]

1

Department of Information Management, Chaoyang University of Technology, Taichung, Taiwan

2

Faculty of Information Technology, Satya Wacana Christian University, Salatiga, Central Java, Indonesia

3

School of Creative Technologies, The University of Portsmouth, Portsmouth, UK

Multimedia Tools and Applications

Systems (ADS) [27] and traffic surveillance. Detecting and recognizing traffic signs are challenging tasks due to several issues, including rotation, occlusion, color variation, and skewing from camera setup in the environment. Moreover, an image with varying sizes, shapes, and colors may contain multiple signs [43, 46]. The arrival of modern breakthroughs in deep learning [36] has significant state-of-the-art results for recognition tasks and object detection [14, 25, 32]. A lot of the research focused on designing deep convolutional neural networks to improve accuracy [17]. Nevertheless, the development of a steady real-time Traffic Sign Recognition (TSR) still presents a challenging problem based on the latency in its testing time, which is crucial in making decisions based on the environment and real-world variability. TSR is one of the most well-known and widely discussed by lots of researchers. The primary