摘要
进行人体运动姿态跟踪时,由于受到光照变化、背景干扰等因素的影响,使得轮廓特征提取不准确,导致跟踪精度降低,对此,本次研究基于轮廓树法的人体运动姿态跟踪方法。首先,采用深度摄像机实现对人体运动姿态高分辨实时监控图像采集,并基于高自由度的肢体关节估计方法,构建人体运动姿态视频图像模型。然后,利用轮廓树特征检测方法,计算分支轮廓质点,提取实时监控视频图像的边缘轮廓特征量。并将根据上述得到的人体运动姿态特征信息,输入到卷积神经网络中进行姿态估计。最后,基于人体姿态估计结果提供的关节位置信息,在连续帧之间进行特征匹配,以实现人体运动姿态跟踪。仿真结果表明,采用本方法对原始图像进行预处理后,明显增加了原始图像的清晰度,并且能够有效实现图像的特征提取,最高跟踪精度达到了98%,识别时间最长仅为49 ms,均优于对比方法,具有一定应用价值。
When tracking human motion posture,due to factors such as changes in lighting and background interference,the extraction of contour features is inaccurate,resulting in a decrease in tracking accuracy.Therefore,this study proposes a human motion posture tracking method based on contour tree method.Firstly,a depth camera is used to capture high-resolution real-time monitoring images of human motion posture,and a human motion posture video image model is constructed based on a high degree of freedom limb joint estimation method.Then,using the contour tree feature detection method,calculate the branch contour particles and extract the edge contour features of real-time monitoring video images.And the obtained human motion posture feature information will be input into the convolutional neural network for posture estimation.Finally,based on the joint position information provided by the human posture estimation results,feature matching is performed between consecutive frames to achieve human motion posture tracking.The simulation results show that after using the method proposed in this paper to preprocess the original image,the clarity of the original image is significantly increased,and the feature extraction of the image can be effectively achieved.The maximum tracking accuracy reaches 98%,and the recognition time is only 49ms,which is better than the comparison method and has certain application value.
作者
李凌君
LI lingjun(Shaanxi Police College,Xi’an 710021,China)
出处
《自动化与仪器仪表》
2024年第10期56-59,64,共5页
Automation & Instrumentation
基金
2022年度陕西省法学课题研究(2022HZ1013)。
关键词
高分辨
实时监控
视频图像
人体运动姿态
跟踪识别
high resolution
real time monitoring
video images
human body movement posture
tracking identification
