An Evaluation of a CNN-Based Parking Detection System with Webcams

In this paper, we evaluate an image processing based parking detection system utilizing convolutional neural networks (CNNs). At present, usage surveys on outdoor parking lots are often performed manually, which may cost a lot. By using commodity webcams and image processing, it may be possible to deploy a parking detection system at a quite low cost. Some parking detection methods utilize HOG and SIFT feature values, and temporal changes of RGB and HSV values. However, these approaches have difficulties due to the influence of ambient light. To tackle this issue, we propose a parking detection method utilizing CNNs, which have high potential in classification and object recognition applications. By training CNNs with different ambient light and lighting conditions, it is expected that the proposed approach can overcome the issue related to the ambient light changes. We evaluate the accuracy of the proposed parking detection system comparing with a method without machine learning, that is, a color-based approach. Experimental results show that the proposed approach can achieve 99 % accuracy for parking and vacancy detection, resulting in an F value of 0.996.