Fully automated on-street parking spot detection with different deep learning methods

Searching for parking spots in metropolitan areas takes a long time, which leads to increase the risk of car accidents, carbon dioxide, and greenhouse emissions. In this dissertation, we propose a vision-based mobile cloud parking management solution, which is fully automated such that it can find roadside parking spots in any street with flowing traffic. We employ both object detection and road segmentation methods to develop an automated parking spot detection system. Thus, we do not need to manually label and train for every distinct street, and do not mark out parking spot boundaries and surrounding road areas. In our approach, the FCN-VGG16 model and KITTI road dataset are used for road segmentation, whereas Faster R-CNN and COCO dataset for object detection. Our Road Boundaries algorithm automatically identifies the road polygon excluding the roadside. The main contribution here is to differentiate the parked cars from the moving cars on the road and detect the available parking spot between the parked cars on the roadside and direct the driver to the nearest spot. On GPU, we achieved a frame rate of 1.5fps and up to 83 percent accuracy with flowing traffic and 92 percent with no traffic flow. These results promise a potential solution on a city-wide scale. Furthermore, three modern convolutional object detection architectures, which are Faster R-CNN, Mask R-CNN, and SSD, were evaluated to select the most optimal running time per frame and detection accuracy trade-offs for parking spot detection. SSD is fastest and Mask R-CNN is slowest in terms of average running time per frame compared to other models.Mask R-CNN has the most detection accuracy and SSD has the least accurate compared to other object detection models.Consequently, Faster R-CNN outperformed in terms of running time per frame and detection accuracy for real time application of parking spot detection.