Deep convolutional neural network for enhancing traffic sign recognition developed on Yolo V4

• Published in: Multimedia tools and applications Vol. 81; no. 26; pp. 37821 - 37845
• Main Authors: Dewi, Christine, Chen, Rung-Ching, Jiang, Xiaoyi, Yu, Hui
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.11.2022; Springer Nature B.V
• Subjects: Advanced driver assistance systems; Artificial neural networks; Computer Communication Networks; Computer networks; Computer Science; Data Structures and Information Theory; Feature extraction; Floating point arithmetic; Intelligent transportation systems; Intelligent vehicles; Multimedia Information Systems; Neural networks; Object recognition; Special Purpose and Application-Based Systems; Traffic control; Traffic information; Traffic signs; Training; CNN; Yolo V4; Object recognition; Yolo V3; Spatial pyramid pooling
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-022-12962-5

Traffic sign detection (TSD) is a key issue for smart vehicles. Traffic sign recognition (TSR) contributes beneficial information, including directions and alerts for advanced driver assistance systems (ADAS) and Cooperative Intelligent Transport Systems (CITS). Traffic signs are tough to detect in practical autonomous driving scenes using an extremely accurate real-time approach. Object detection methods such as Yolo V4 and Yolo V4-tiny consolidated with Spatial Pyramid Pooling (SPP) are analyzed in this paper. This work evaluates the importance of the SPP principle in boosting the performance of Yolo V4 and Yolo V4-tiny backbone networks in extracting features and learning object features more effectively. Both models are measured and compared with crucial measurement parameters, including mean average precision ( mAP ), working area size, detection time, and billion floating-point number (BFLOPS). Experiments show that Yolo V4_1 (with SPP) outperforms the state-of-the-art schemes, achieving 99.4% accuracy in our experiments, along with the best total BFLOPS (127.26) and mAP (99.32%). In contrast with earlier studies, the Yolo V3 SPP training process only receives 98.99% accuracy for mAP with IoU 90.09. The training mAP rises by 0.44% with Yolo V4_1 ( mAP  99.32%) in our experiment. Further, SPP can enhance the achievement of all models in the experiment.