A one-stage anchor-free keypoints detection model for fast electric vehicle charging port detection and pose extraction.

Abstract

As intelligent technologies advance in electric vehicles (EVs), automatic unmanned charging systems are becoming increasingly prevalent. A key breakthrough lies in developing efficient methods to identify and locate charging ports. However, challenges such as high sensor costs, compromised robustness in complex environments, and stringent computational demands remain. To address these issues, this study introduces FasterEVPoints, a state-of-the-art convolutional neural network (CNN) model integrating partial convolution (PConv) with FasterNet. Tailored to pinpoint critical points of EV charging ports, FasterEVPoints incorporates the perspective-n-point (PnP) algorithm for pose extraction and the bundle adjustment (BA) optimization algorithm for refined pose accuracy. This approach operates effectively with only a single RGB camera, ensuring precise localization with minimal hardware. Experiments demonstrate that in complex lighting scenarios, FasterEVPoints boasts 95% detection accuracy on a proprietary dataset with a positioning error of less than 2 cm at a 50 cm distance. Furthermore, when integrated into the you only look once X (YOLOX) framework with parameters comparable to YOLOX-Tiny, FasterEVPoints delivers similar accuracy while consuming only 73% of the computational load and 66% of the parameters compared to YOLOX-Tiny. This exceptional efficiency, combined with high detection accuracy, establishes FasterEVPoints as a practical and scalable solution for real-world autonomous EV charging applications.