Generating Parking Area Patterns from Vehicle Positions in an Aerial Image Using Mask R-CNN

Abstract

Identification of visual patterns in cluster of similar objects in an aerial scene is a challenge for drone surveillance. Similar objects may form different patterns representing non-identical meanings in a scene. For example, a group of cars in a scene may form either traffic or a parking area, subjected to their position and arrangements. In this work, we proposed an approach to this issue by distinguishing between a parking area and any other arrangement of cars by generating a grouping pattern. Initial object detection in the aerial images is carried out by Mask-RCNN and then an algorithm is developed to form the arrangement pattern of the vehicles. A comparative analysis between parking spaces and non-parking spaces is provided in this work and distinct variations were reported between the generated patterns. The pattern generated showed formation of distinct geometric shapes in case of cars in a parking plot whereas cars in a traffic or non-parked condition displayed uneven shapes where many edges intersected at various points. However, the pattern generated will depend on the elevation of the camera from which the scene is captured and the ability of the object detection algorithm to detect cars. This approach can be used for understanding parking area organization, traffic managements, and direct cluster identification using deep learning methods for aerial surveillance applications in the near future.