Abstract:
Vision-based surround-view free space detection is crucial for automatic parking assist. In this task, precise boundary localization is the most concerned problem. In this paper, we have proposed to reframe the free space as polar representation for the free space boundary, and exploit a transformer framework to regress the representation end-to-end. To restrain the overall shape of the free space, we have introduced a Triangle-IoU loss function, enabling the network to consider the boundary as a whole. Furthermore, we have proposed a challenging newly-built surround-view dataset (SVB) with boundary annotations and supplied a new metric for boundary quality. Experiments on SVB dataset validate the effectiveness of our method, which outperforms existing free space detection methods and runs in real-time with a remarkable reduction in the computational cost. Additionally, our method shows excellent generalization ability to new parking scenes.
