Detect the Video Recording Act of UAV through Spectrum Recognition

Spectrum recognition technology could be an important issue to monitor the video recording act of UAV. The existing methods cannot recognize the UAV spectrum if it coexists with other spectra in the same time-frequency map. This paper proposes Safe-UAV, a complex semantic segmentation model, to solve the problem. Safe-UAV utilizes the time-frequency waterfall map(TWM) as input. We adopt a feature extraction module to obtain the convolution of TWM with the different receptive fields. We also propose a feature fusion unit to implement a grading fusion strategy according to the depth of different convolution layers. Once the training process is complete, Safe-UAV can recognize every pixel of the input TWM. Moreover, we collected ten types of civilian UAVs and built a dataset to evaluate the model performance. The experimental results show that our model ensures high recognition accuracy without reducing the efficiency compared with other algorithms such as espnet and fcn8s.