基于神经辐射场的无人机拼接图像遮挡目标解耦检测系统设计

摘要:

在强电磁干扰与多径散射耦合的工业复杂电磁环境中,采集的宽幅拼接影像易因微镜阵列单元失配与高频振荡引入的相位同步误差,形成非均匀辐照伪影与动态像移,导致小样本训练的标签缺失,增加检测难度。对此,设计了基于神经辐射场的无人机拼接图像遮挡目标解耦检测系统。通过无人机与地面端的协同硬件及多GPU服务器设计硬件模块,保障了数据处理流程的高效性。融合符号距离场与双向反射分布函数,并引入体密度修正,实现了优越的三维隐式重建,通过正/负检测头分离前景与背景,并利用带多标签约束的损失函数,有效克服了非均匀辐照伪影与动态像移环境下,小样本训练的标签缺失问题,达成了精准的任务解耦。在高反射过曝遮挡场景下,本系统在特征点匹配上实现零丢失、零误匹配,边界框最大误差频率降低至8次,验证了其在复杂遮挡环境下的有效性。

Abstract:

In the complex industrial electromagnetic environment where strong electromagnetic interference and multipath scattering are coupled, the wide swath spliced images collected are prone to non-uniform irradiation artifacts and dynamic image shifts due to phase synchronization errors introduced by mismatch of micro mirror array units and high-frequency oscillations, resulting in label loss during small sample training and increasing detection difficulty. A neural radiation field based unmanned aerial vehicle (UAV) image stitching occlusion target decoupling detection system was designed for this purpose. The efficiency of the data processing process is ensured through the collaborative hardware between the drone and the ground, as well as the design of hardware modules with multiple GPU servers. By integrating the symbol distance field and bidirectional reflection distribution function, and introducing volume density correction, superior 3D implicit reconstruction has been achieved. The foreground and background are separated by positive/negative detection heads, and a loss function with multi label constraints is used to effectively overcome the problem of label loss in small sample training under non-uniform irradiation artifacts and dynamic image shift environments, achieving accurate task decoupling. In high reflection overexposure occlusion scenes, this system achieves zero loss and zero error matching in feature point matching, and reduces the maximum error frequency of bounding boxes to 8 times, verifying its effectiveness in complex occlusion environments.