【摘要】 在地基实况测量望远镜中，通常包含捕获和成像两组光学系统。望远镜使用捕获系统对目标闭环跟踪时，由于视场较大，跟踪精度低，可明显感受成像系统的视频抖动，不利于目标实时观察和相关科学实验。为解决上述问题，本文提出一种支持场景切换感知的电子稳像技术。首先，目标未出现时，通过最大连通区域法对每帧图像尝试提取目标特征区域，直至成功提取。接着，对于每帧图像，使用核相关滤波器对上一帧训练模型进行特征匹配，并更新训练模型。其次，使用最小二乘法拟合目标运动轨迹，通过误差分析判断目标真伪。最后，根据目标真伪判断结果，决定当前图像的裁剪是否使用抖动补偿。实验结果表明：经过该技术处理的视频，目标稳定性具有较大提升，相邻图像特征区域抖动幅度最大值从±10 pixels降低至约±1 pixel，使用前50帧图像作为参考帧的峰值信噪比平均提升4.62 dB；同时算法能够感知目标进入和离开视场；算法处理时间小于2 ms，满足实时处理需求。实现连续观测多目标的稳像流程自动化功能，提升相关科学实验的质量和成功率。更多还原

【Abstract】 Ground based real-time measurement telescopes typically include two sets of optical systems for capture and imaging. When a telescope uses a capture system to track an object in a closed-loop manner, due to the large field of view and low tracking accuracy, the video jitter of the imaging system can be clearly felt, which is not conducive to real-time observation of the object and related scientific experiments. To address the aforementioned issues, this article proposed an electronic image stabilization technology that supports scene switching perception. Firstly, when the object did not appear, the maximum connected region method was used to attempt to extract the object feature region for each frame until it was successfully extracted. Then, for each image, a kernel correlation filter was used to match the features of the previous frame’s training model, and the training model was updated. Next, the least squares method was used to fit the object motion trajectory and determine the authenticity of the object through error analysis. Finally, determined whether to use jitter compensation for cropping the current image based on the authenticity of the object. Experimental results indicate that the video processed by this technology has a significant im-provement in object stability, the maximum jitter amplitude of adjacent image feature regions decreases from ± 10 pixels to approximately ± 1 pixel, the peak signal-to-noise ratio of reference frames within 50frames is increased by an average of 4. 62 dB; Simultaneously, the algorithm can perceive the object entering and leaving the field of view; The algorithm processing time is less than 2 ms, meeting the online processing requirements. Realize the automation of the image stabilization process for continuous observation of multiple objects, and improve the quality and success rate of related scientific experiments.更多还原