Implementation of CNN-based parking slot type classification using around view images

This paper presents a commercial implementation of CNN-based classification of parking slot type using around view images. The existing automatic parking systems use ultrasonic sensors, but they often fail to classify the types of parking slots. Around view images can depict the types of parking slots distinguishably. However, due to the diverse lighting and ground conditions, it is difficult to classify the parking slot type using images. Moreover, it is hard to find the parking lines since the lines are often occluded by vehicle or erased. To overcome these problems, we have constructed an extensive dataset composed of labeled 480,000 images acquired on various environments using around view monitoring (AVM) camera mounted on a commercial vehicle. For training our CNN-based classifier, we subdivided parking slot types into ten categories and finally derived three parking slot types for actual applications. To operate the classifier in real vehicles, we designed CNN model suitable to embedded systems and implemented it using GPU. In experimental evaluation, the implemented CNN-based classifier achieves an accuracy of 94.15%. And the processing time of our classifier achieves 3.67ms per frame (270 fps) on the NVIDIA Tegra CX embedded system.