Recognition of Engineering Vehicles in Aerial Images of Multi Rotor UAV

It is one of the significant tasks of power inspection by multi rotor Unmanned Aerial Vehicle (UAV) to recognize engineering vehicles in aerial images. If there are engineering vehicles working near or below the high-voltage power line, the UAV would give out the important early warning message to avoid the situation that the bucket or boom of the engineering vehicle enters the safe distance from the high-voltage power line, and reduce accidents such as short circuit breakdown. Aiming at the problem of recognition of engineering vehicles in aerial images of UAV inspection, this paper proposed an improved capsule network method. First, the structure of original capsule network is replaced with a multi-layer densely connected capsule network. Next, the dynamic routing algorithm of the capsule network is improved. As the results of experiments have shown, (1) the improved capsule network method gets a mAP of 93.74% for the recognition of engineering vehicles, and its parameter scale is smaller than other methods. (2) The number of network layers influences the recognition precision greatly. Their relationship is non-monotonic and nonlinear. In addition, whether or not to improve the dynamic routing algorithm does not affect the trends of recognition mAP. The overall performance of the improved capsule network method is obviously better than YOLOv5 and other artificial feature extraction methods.