Identification of Safe Landing Areas with Semantic Segmentation and Contour Detection for Delivery UAV

Abstract

Automatic identification of landing area helps UAV (Unmanned Aerial Vehicle) to land correctly and safely. UAV can use computer vision to detect special signs, such as QR code, April tag, and geometric shapes. However, such object detection lacks the ability to fully recognize its surrounding environment. Currently, advances in deep learning have the potential to replace these special markers and can recognize other objects around the landing area more accurately. Moreover, delivery UAV operating in urban areas are prone to uncertain conditions. Thus, this paper proposes a semantic segmentation method to segment objects in the delivery UAV landing area and identify safe landing area. Contour detection is used to extract the segmentation results to get the largest polygon in the contour and coordinate points of the landing target. The process of predicting each image takes 0.52 seconds. This overall method was successfully executed by using an i3 processor with an average processing speed of 4 frames per second. The results showed that the model has a good segmentation ability to identify the target of UAV landing area.