Fast Recognition and Localization of Electric Vehicle Charging Socket Based on Deep Learning and Affine Correction

With the popularity and intelligence of electric vehicle, the increasing demand for charging convenience has driven the development of automatic charging technology. The recognition and localization of electric vehicle charging socket is the key to automatic charging. This study proposes a system for fast recognition and localization of electric vehicle charging socket based on deep learning and affine correction. First, modify the yolov4 network structure for recognizing the charging socket to improve the recognition speed. Second, using the meanshift clustering algorithm, the noise is effectively removed to improve the recognition success rate. Third, we propose a pixel coordinate correction method for the charging socket based on the affine transformation. The projective transformation is approximated to the affine transformation when the camera is facing the charging socket. According to the properties of covariance and distance ratio invariance, the pixel coordinates of the charging holes are corrected. Finally, the charging socket is located by the Perspective-n-Point (PnP) algorithm. With different angles, distances and light intensities, the recognition success rate of the charging socket is 100%, and the average recognition time for single-frame image is 27ms. The localization accuracy is tested under different light intensity and distances. After affine correction, the localization accuracy is improved, and the final average localization errors are 1.418 degrees, 1.660 degrees, 0.050 degrees, 0.217mm, 0.215mm and 0.855mm in Rx, Ry, Rz, x, y and $z$ respectively. The results show that our method has a good effect on the recognition and localization of the charging socket in complex environment.