| 引用本文: | 柳恩涵,张锐,赵硕,王茹.一种基于视频预测的红外行人目标跟踪方法[J].哈尔滨工业大学学报,2020,52(10):192.DOI:10.11918/201910048 |
| LIU Enhan,ZHANG Rui,ZHAO Shuo,WANG Ru.Infrared pedestrian target tracking method based on video prediction[J].Journal of Harbin Institute of Technology,2020,52(10):192.DOI:10.11918/201910048 |
|
| 摘要: |
| 红外视频与普通彩色视频相比易受周围环境的影响,在红外行人跟踪中行人目标外观轮廓和灰度分布常有较大幅度变化,导致跟踪困难.为解决此问题,本文提出了一种VPSiamRPN（Video Prediction with Siamese Region Proposal Network）红外行人目标跟踪系统.在SiamRPN（Siamese Region Proposal Network）网络的基础上,针对目标形变、目标遮挡和背景杂波等严重影响红外跟踪效果的因素进行研究,设计将PredNet (Deep Predictive Coding Networks for Video Prediction and Unsupervised )网络的图像预测功能结合应用到SiamRPN网络上,以提高跟踪模板与被检测目标的相似度,增强目标跟踪中的模型匹配能力,从而提高对红外行人目标的跟踪能力.通过改变网络层数、预测所用的目标图像及图像帧数、网络的跟踪策略,对网络进行优化,设计了9组对比试验,在PTB-TIR数据集上,与SiamRPN网络客观定量对比,通过跟踪的成功率和重叠率在10种属性上对网络进行评估.实验结果表明:本文网络对红外目标的识别在热交叉、强度变化、遮挡和尺寸变化等多种属性上的跟踪成功率和重叠率均较SiamRPN网络有较大提高,显示出对红外行人跟踪的良好性能,在这一领域将有广阔的应用前景. |
| 关键词: 目标跟踪 视频预测 孪生网络 红外 PredNet |
| DOI：10.11918/201910048 |
| 分类号:TP391 |
| 文献标识码:A |
| 基金项目: |
|
| Infrared pedestrian target tracking method based on video prediction |
|
LIU Enhan1,ZHANG Rui1,ZHAO Shuo2,WANG Ru1
|
|
(1.Collage of Automation, Harbin University of Science and Technology, Harbin 150080, China; 2.Collage of Computer Science and Technology, Harbin Engineering University, Harbin 150001, China)
|
| Abstract: |
| Compared with the common color video, infrared video is easily affected by the surrounding environment. In infrared pedestrian tracking, the appearance contour and gray distribution of the pedestrian target often have great changes, which lead to the difficulty of tracking. To solve this problem, this paper proposes a VPSiamRPN (Video prediction with Siamese Region Proposal Network) infrared pedestrian target tracking system. Aiming at the factors that seriously affect the performance of infrared pedestrian tracking (such as target deformation, target occlusion, and background clutter), the image prediction function of PredNet (Deep Predictive Coding Networks for Video Prediction and Unsupervised) was designed and applied to SiamRPN (Siamese Region Proposal Network) to improve the similarity between the tracking template and the detected target, so as to improve the tracking ability to the infrared pedestrian target. Nine comparative experiments were carried out by changing the number of layers of the network, the number of target images and frames used for prediction, and the tracking strategy of the network. On PTB-TIR dataset, experimental results show that the success plots and precision of theinfrared target recognition in thermal crossover, intensity change, occlusion, scale variation, and other attributes were much higher than those of SiamRPN, indicating good performance of infrared pedestrian tracking, which will have broad application prospects in this field. |
| Key words: target tracking video prediction Siamese network IR image PredNet |
