基于多特征分析的井下视频关键帧提取及人员检测

Key frame extraction based on multi-feature analysis and personnel detection of downhole video

针对井下视频人员检测准确度和效率低的问题,提出了一种基于多特征分析的视频关键帧提取技术和YOLOX算法检测相结合的方法。该方法对井下视频进行关键帧提取处理,使用灰度直方图的欧式距离、HSV颜色空间量化后的余弦距离、方向梯度直方图的相似性和选择结构相似性来计算相邻帧之间的特征相似度。通过动态权重参数的方法融和以上指标作为最终的相似度数值,并与设定的阈值相比较来提取视频的关键帧集合。将提取的关键帧集合在训练好的YOLOX模型中进行检测。通过不断地实验和改进,结果表明,使用该方法对井下视频提取的关键帧冗余度小、分布均匀,能够全面地表达视频的内容。在人员检测方面,能够以更快的速度和更高的精确度实现对井下人员的检测。

Aiming at the problem of low detection accuracy and efficiency of underground video personnel,a method combining video key frame extraction technology based on multi-feature analysis and YOLOX algorithm detection is proposed.This method extracts key frames from the downhole video,and uses the Euclidean distance of the gray histogram,the cosine distance after quantization of the HSV color space,the similarity of the direction gradient histogram and the structural similarity to calculate feature similarity between adjacent frames,and through the method of dynamic weight parameters,the above indicators are fused as the final similarity value,and compared with the set threshold to extract the key frame set of the video,finally,the extracted keyframe set is tested in the trained YOLOX model.Through continuous experiment and improvement,the results show that the key frames extracted from downhole video using this method have small redundancy and uniform distribution,and can fully express the content of the video.In terms of personnel detection,the detection of underground personnel can be realized with a greater speed and higher accuracy.