eMNv4-YOLO: A High-Efficiency Target Detection Framework for Robotic Driving Vehicle Instrument Reading
Yin Lei, Shan Chen, GE Wan-cheng, Xinhong Wang, Jianyao Hu · IEEE Access · 2025
In the domain of autonomous vehicle testing, a crucial challenge in implementing onboard information perception systems is the development of lightweight detection network models that can be efficiently deployed on edge devices. In this study, we collected and collated 6,168 images encompassing automotive dashboards, industrial instruments, and automotive simulation test instruments, which possess diverse features. Subsequently, we performed manual annotation on these images to construct the instrument detection dataset, namely Complex_Dashboard_Dataset. To satisfy the environmental perception requirements of mobile driving robots, this paper presents an enhanced feature recognition and detection model for automotive dashboards, named eMNv4-YOLO.This model combines the YOLO object detection framework with the lightweight deep learning model MobileNetV4. Moreover, a novel lightweight convolutional module, EPDConv, was introduced in this study to further reduce the model parameters, thereby yielding the refined eMNv4 model. In addition, a highly robust and efficient automotive dashboard reading recognition algorithm, SAGR, was developed by integrating machine vision image processing algorithms and template matching algorithms. The experimental results show that the eMNv4-YOLO model achieved an mAP50 detection result of 96.4% on the constructed dataset, whereas the model parameter count was reduced to only 2.0M, substantially reducing the model size and computational complexity. Finally, the verification of a real vehicle validated the effectiveness of the SAGR algorithm. In conclusion, the eMNv4-YOLO model designed in this study has extensive applicability to information perception systems on edge mobile devices, providing an effective object detection solution in the field of autonomous driving.