Semantic Map Construction of UAV Autonomous Landing in Unknown Environment

Aiming at the problems of environment identification and dynamic target information detection in the process of autonomous landing map construction of quadrotor UAVs in unknown environments, a method of autonomous landing semantic map construction of UAVs in unknown environments is proposed. Firstly, an improved Deeplabv3+ model is proposed. The image collected by UAV is restored based on the Lucy Richardson algorithm. On this basis, the information after fusion processing of the collected original image features and restored image features is input into the semantic segmentation model. Shufflenetv2 is used as the backbone network of Deeplabv3+ and the CBAM attention mechanism is introduced to improve the accuracy of the semantic segmentation model. Secondly, to improve the accuracy of dynamic target detection, a dynamic target detection method based on optical flow and semantics is designed. The environmental semantic information is introduced into the Lucas Kanade algorithm to estimate the motion displacement of the airborne camera to determine the optical flow threshold of the dynamic target. Then, a semantic grid map is established to model the environment, define the measurement index of the safe landing sites, and design the selection strategy of the safe landing sites. The ICG Drone Dataset is used for experiments, and the results show that the image segmentation accuracy of the designed method is improved by 4.3% on average compared with the traditional DeeplabV3+ model, and the detection rate of the dynamic targets reaches 93.8% correctly. The measured results show that the drone can accurately detect the dynamic targets in an unknown environment and autonomously select the safe area as the landing site.