基于Transformer模块和CNN的无人机避障方法研究

Research on Obstacle Avoidance Method of UAV Based on Transformer Module and CNN

针对传统CNN避障方法无法获得全局感受野、图像特征提取计算量大的问题,以四旋翼无人机为研究对象,提出一种基于Swin Transformer模块改进CNN模型的无人机避障方法。首先,使用Swin Transformer代替CNN模型中的Conv2D层,进行全局信息特征提取;然后,构建3个残差结构相连的Swin Transformer网络,输出无人机在当前飞行环境下的转向预测和碰撞预测;最后,设计无人机多姿态映射控制系统,输出无人机避障控制指令。实验结果表明,所提方法碰撞预测平均准确率为96.8%,转向预测均方根误差(RMSE)为0.068,满足了无人机自主避障的要求。

Aiming at the problems of the traditional CNN obstacle avoidance method,which can not obtain the global receptive field and the large amount of calculation for image feature extraction,taking the quadrotor UAV as the research object,a UAV obstacle avoidance method based on Swin Transformer module to improve the CNN model is proposed.Firstly,Swin Transformer is used to replace the Conv2D layer in the CNN model to extract global information features.Then,the Swin Transformer network connected with three residual structures is constructed to output the steering prediction and collision prediction of the UAV in the current flight environment.Finally,the UAV multi-attitude mapping control system is designed to output the UAV obstacle avoidance control command.Experimental results show that the average accuracy of collision prediction was 96.8%,and the root mean square error(RMSE)of steering prediction was 0.068,which meets the requirements of autonomous obstacle avoidance of UAV.