基于多轴自注意力的无人机避障模型

UAV Obstacle Avoidance Model Based on Multi-axis Self-attention

针对无人机在飞行过程中容易因旋翼碰撞而坠毁的问题,提出利用改进的图像识别模型实现自动预警。将瓶颈多轴自注意力模块(BMSA)嵌入到图像识别模型中进行改进,提升模型对细小物体的识别准确率。多轴自注意力层在低分辨率阶段替换原本卷积层,使得模型能够兼顾局部自注意力和全局自注意力。实验结果表明:改进得到的多轴自注意力的残差网络(MS-ResNet)具有较高的障碍物识别准确率,能实现较好的预警效果。

To address the proneness of UAV crash due to rotor collision during flight,an improved image recognition model is proposed to achieve automatic warning.A bottleneck multi-axis self-attention module(BMSA)is embedded into the image recognition model for improvement,enabling the model to improve the recognition accuracy of the model for fine objects.The multi-axis self-attentive layer replaces the original convolutional layer in the low-resolution stage,enabling the model to obtain both local self-attention and global self-attention.The experiments show that the improved multi-axis self-attentive residual network(MS-ResNet)has high accuracy of obstacle recognition and achieve a better early warning effect.