YOLO-based lightweight traffic sign detection algorithm and Mobile deployment

Abstract

In response to the problems of large parameter quantity, poor real-time performance and low accuracy in small targets and complex background traffic sign detection, based on the self-made road traffic sign dataset, this paper proposes a lightweight traffic sign detection algorithm based on Yolo and implements mobile deployment. Firstly, Utilizing the C2f lightweight network architecture, we employ variable convolution and BiFPN_Concat operations to augment the adaptability of the backbone network towards geometric transformations across objects of varying sizes; Secondly, After the C2f layer in the backbone network, the SimAm attention mechanism is introduced to prioritize key features of small-sized targets and reduce model complexity, facilitating the development of a lightweight detection system; Next, the Focal EIoU loss function is integrated to adjust the weights of challenging samples, addressing the issue of sample imbalance during classification; Finally, using Android Studio platform, we accomplish the comprehensive interface design and functional implementation of the mobile app. The deployed mobile app achieves real-time and accurate detection of traffic signs under real-world road conditions, even in offline mode. The experiment also conducted a vertical comparison between the SimAm and EMA attention mechanisms at different network levels to determine the optimal level and position for embedding the attention mechanism. Additionally, the algorithm proposed was horizontally compared with mainstream algorithms, and the improved method was evaluated on a custom dataset. The obtained metrics, including F1 score, mAP@0.5, and mAP@0.5:0.95, are 0.8987, 98.8%, and 75.6% respectively. The detection speed achieves 50 FPS, meeting the requirements for real-time detection.