Abstract：To solve the problem regarding parking slot detection during automated valet parking (AVP) process， a hybrid encode-decode structured network model was presented using improved Transformer module and convolution. “T” and “L” shaped parking slot corner points were detected with above mentioned network and position， direction and type information of corner points were obtained by regression. Then， corner points were put into an inference module which could detect parking slot base on types and geometric relations between corner point pairs. Finally， multiple experiments regarding a variety of scenarios were performed on ps2.0 parking slot dataset， with precision rate and recall rate of detection reaching 99.36% and 99.31%， respectively，while RMSE of position error was only 0.87 pixel. The computational cost to detect corner points of a single image was around 2.71 GFLOPs and corresponding detection time was about 40 ms.
2024, Vol. 58 
