基于深度学习的认知成像激光雷达(特邀)

Cognitive Imaging Lidar Based on Deep Learning(Invited)

为提升传统成像激光雷达系统对场景的感知能力和信号处理算法的泛化能力,提出了一种基于深度学习的认知成像激光雷达方法,通过深度学习点云目标检测算法的处理结果进一步调控核心成像参数,形成认知反馈,提升系统成像质量和环境感知能力。为验证该方法的可行性,设计并实现了一套认知成像激光雷达演示模块,通过实验对比分析,选择激光器的发射功率、成像系统的扫描视场和扫描角分辨率三个成像参数进行认知反馈,并结合深度学习方法实现了与场景的动态交互学习,解决了传统激光雷达成像参数固化的问题。实验结果表明,采用基于深度学习的认知成像工作模式有效提升了现有深度学习点云目标检测算法的泛化能力和目标检测精度。

To improve the scene perception ability of traditional imaging lidar system and the generalization ability of the signal processing algorithm,a cognitive method of imaging laser radar based on deep learning is proposed.Through the result of deep learning point cloud target detection algorithm,the core imaging parameters are further regulated,and the cognitive feedback is formed,improving the system imaging quality and environmental perception.To test and verify the feasibility of the proposed method,a cognitive imaging laser radar presentation module is designed and implemented.Through the experimental comparison and analysis,three imaging parameters of laser emission power,scanning field of view and scanning angular resolution of imaging system are selected for cognitive feedback,and the module combined with the deep learning method realize the dynamic interaction learning of the scene,which has solved the problem of solidification of traditional lidar imaging parameters.The experimental results show that the cognitive imaging mode based on deep learning can effectively improve the generalization ability and target detection accuracy of the existing deep learning point cloud target detection algorithm.