KEPS-NET: Robust Parking slot Detection based Keypoint estimation for High Localization Accuracy

In this paper, we study parking slot detection problem for automated parking function. Main contribution is to propose a neural network to detect parking slot based on keypoint estimation, which is called KEPS-NET. The proposed network detects three kinds of parking slot (perpendicular, parallel, and slanted) with two entrance corner points. We first analyze how precise localization of the entrance points in pixel level is important for real automated parking scenario. Then we construct parking slot detector with four heads using a common feature extraction backbone network to estimate center point, entrance points, slot type, and slot angle. And separate key-point detector with two heads is added to find entrance points more accurately by leveraging richer local features. We also design a slope loss helping the network minimize the orientation errors of the entrance line, which allows the network to obtain greatly improved localization accuracy. Through the experimental results, we show that the proposed algorithm achieves state-of-the-art performance compared to the previous works.