Deep Learning based Parking Spot Detection and Classification in Fish-Eye Images

Within the universe of automated driving (AD) applications, automated valet parking (AVP) is especially attractive in terms of opportunities and adoption. A camera is one of the commonly favored perception sensors in the AVP application. Where the detection and classification of the parking spot as free or occupied in the camera feed are critical. The parking spot detection functionality is used during an autonomous operation to locate and maneuver into the appropriate parking location. The main challenge in the detection of a parking spot lies in the manifestation of parking spots as trapezoid shapes in the image domain. Further, the contemporary cameras used in surround view or park assist features tend to be fish-eye, where lens distortion causes real-life straight line segments and objects appear curved. This paper presents an efficient, pixel level parking spot instantiation and classification approach based on object detection framework in deep learning. The approach uses MobileNet-V1 network architecture as backbone convolutional neural network (CNN) and modified Single Shot Detector(SSD) as object detection meta-architecture to performs parking spot instantiation directly in the fish-eye domain. The conventional SSD meta-architecture of object detection functionality has been innovatively augmented to detect major key-points of the parking spots to delineate parking spot boundary precisely. The proposed solution achieves 0.87 mAP for detected rectangular boxes enclosing parking spots and accuracy of 0.76 OKS in four corner points per detection.