Abstract: Low altitude airspace drones are characterized by small size and flexible flight, which brings difficulties to visual detection of trespassing drones. The low altitude drone object segmentation algorithm incorporating an attention mechanism named Rep YOLACT (re-parameterization you only look at coefficients network) is proposed, which is first used with RepVGG (re-parameterization visual geometry group) networks to improve ResNet (residual network) backbone in YOLACT and enhance the feature extraction capability of the network. Meanwhile, CBAM (convolutional block attention module) is added after the three feature layers output from the backbone feature extraction network, so as to further utilize the information of the feature layers efficiently. Experiments are conducted on FL-drones (flying drones dataset) and MUD (multiscale unmanned aerial vehicle dataset), respectively. The results show that the proposed Rep YOLACT algorithm improves mask AP (average precision) and mask AR (average recall) by 0.3% and 11.7%, respectively, compared with YOLACT algorithm on FL-drones. The proposed Rep YOLACT algorithm improves 2.3% and 5% on mask AP and prediction frame AR compared to YOLACT algorithm, which can perform the drone segmentation task well and its segmentation accuracy is higher than other mainstream segmentation algorithms.

Key words: drone, object segmentation, attention mechanism, RepVGG network, deep learning