Evaluation of Parking Space Detection From Aerial Imagery Using Convolutional Neural Networks
Abstract: In this thesis, the viability of using Convolutional Neural Networks to detect parking spaces using aerial imagery has been evaluated. Three state of the art networks have been tested - YOLOv3, RetinaNet, and Mask R-CNN. A dataset of urban parking lots and corresponding annotations was generated from scratch using a custom built GUI to annotate automatically generated images of parking lots from Open Street Map, from varying aerial imagery providers. This dataset was used to test and evaluate the different networks, and Mask R-CNN was used for a lengthy parameter tuning process, as it seemed to perform optimally of the three networks. The resulting model did not perform the best, believed to be because of the low amount of features represented in parking spaces. While results indicated that a somewhat complete solution is possible, it might not be feasible using a pure single CNN approach.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)