一种改进MDnet的人物跟踪模型

An improved MDnet character tracking model

在机场、高尔夫球场等场所,设计一种载重式跟踪机器人帮助人们托运重物,能够识别人物并且进行自主跟踪。该识别方法改进了MDnet在线跟踪模型,能在人物被遮挡以后保持原来的跟踪效果。原生的MDnet在线跟踪算法,没有处理目标被遮挡的情况下的图片帧,当目标的遮挡情况严重时跟踪就会丢失。为了解决这个问题,在输入前加入Img⁃Align层,使用双线性内插值方法进行图像尺寸处理,得到更精确的目标值;在全连接层后加入Occlusion⁃Assess层,设置函数变换,降低目标遮挡严重的帧的正样本采集范围,保证之后跟踪的正确率。对3个数据集10760张图片进行训练,改进后的模型交并比提高了10%,成功跟踪评分提升了15.6%,跟踪时间平均减少了24%。

In the airport,golf course and other places,a kind of load⁃bearing tracking robot is designed to help people consign heavy objects,which can recognize and track people independently.This recognition method improves the MDnet online tracking model,and can keep the original tracking effect after the characters are occluded.The original MDnet online tracking algorithm does not deal with the picture frame with the occluded target,and the tracking target will be lost when the target is occluded seriously.In order to solve this problem,the Img⁃Align layer is added before the input,and the bilinear interpolation method is used to process the image size to get more accurate target value;the Occlusion⁃Assess layer is added behind the full connection layer,and the function transformation is set to reduce the sample collection range of the frame with seriously⁃occluded target,so as to ensure the accuracy of the subsequent tracking.After training with 10760 pictures from three datasets,the cross union ratio of the improved model is increased by 10%,its success tracking score rises 15.6%,and its average tracking time is reduced by 24%.