四旋翼无人机视觉目标检测研究

Research on Target Detection of Quadrotor Based on Vision

笔者从四旋翼无人机的自主飞行技术入手,基于计算机视觉技术和TensorFlow框架实现四旋翼无人机目标检测,采用Pixhawk开源飞控,在获取视觉传感器采集的图像数据后,利用TensorFlow深度学习框架搭建快速区域建议网格学习模型进行目标检测标定。经实验验证发现,模型能够实现快速收敛,并实现损失函数保持较低值,在分析不同参数和卷积模型对目标识别产生的影响后,发现在较小目标情况下存在漏检和不稳定现象。

Starting with the autonomous flight technology of quadrotor UAV,the target detection of quadrotor UAV is realized based on computer vision technology and tensorflow framework.Pixhawk open source flight control is used.After obtaining the image data collected by visual sensors,the tensorflow deep learning framework is used to build a fast region recommendation grid learning model for target detection and calibration.The experimental results show that the model can achieve fast convergence and keep the loss function at a low value.After analyzing the influence of different parameters and convolution model on target recognition,it is found that there are missing detection and instability in the case of small targets.