Abstract:
Current mainstream methods adopt point regression and prior sizes to detect parking slots. Although these methods are approved to be useful in majority of circumstances, they have many limitations when they adapt to size and shape variations in more general slots because of pre-defined fixed order, lack of holistic constraints and adequately diverse data. To address them, we propose an order-independent matching strategy with shape similarity to handle the more general slot sizes and shapes. The matching strategy adopts a two-level procedure: the point-level and the parking slot-level, that finds optimal order and association adaptively. More importantly, we adopt shape similarity to represent holistic geometry to rank slots so as to suppress the misshapen ones. Furthermore, we collected a large-scale and remote-view parking slot dataset (LRPS) to improve data diversity. It contains a large number of general parking environments, as well as slots of various shapes and sizes across different cities, daytime and interior-exterior scenes. The proposed approach is evaluated on the LRPS dataset and achieves superior performance to previous methods.
