传感器网络中基于压缩感知的压缩域目标跟踪算法研究与应用

Research and application of target tracking algorithm based on compressed domain in wireless sensor network

复杂环境下传感器网络目标跟踪,存在跟踪准确性与算法复杂性这对矛盾,需要考虑准确性、通信和计算能耗之间的折中。针对此问题,研究传感器网络中基于压缩感知的压缩域目标跟踪和压缩域数据融合,并提出了用稀疏的测量矩阵对haarlike特征压缩,压缩域特征送入朴素贝叶斯分类器,在压缩域直接实现目标跟踪的算法。然后通过配置传感器网络以生成多个层次类型不同的簇结构,压缩后的数据在网络中传输,并在各层簇头实现压缩域下的数据融合。该算法通过稀疏测量矩阵压缩表征原始图像信息和分类器的自我学习更新,提高了对压缩域目标特征分类的准确性,在复杂环境下有更好的鲁棒性。而压缩域直接进行目标跟踪,不需要重构图像,也减少了网络运算量和数据传输量。通过仿真实验和标准数据库测试对比以及在机器人足球赛实验平台中的应用表明,该算法在跟踪准确性,数据传输量及传输能耗上均有优势。

Target tracking of sensor networks in complex environment has contradictions between the tracking accuracy and the complexity of the tracking algorithm. The tradeoff between accuracy and energy consumption of communication and calculation should be considered. To solve this problem,the compressed-domain target tracking and compressed-domain data fusion based on compressive sensing are studied,and a target tracking algorithm performed in compressed domain is proposed. The Haarlike features of original images are compressed by a sparse measurement matrix,fed into the Bayesian Classifier directly in compressed domain. By configuring the sensor networks to generate cluster structures with different hierarchical types,the compressed data is transmitted in the network,and the data fusion in the compressed domain is performed in the cluster heads of each layer. The classification accuracy of target feature in compressed domain is improved by both the sparse measurement matrix and the self-learning of the classifier. The amount of computation and data transmission of networks are reduced by tracking directly in compressed domain without reconstructing. The results of simulation experiment,comparison test of standard database and application in robot soccer game of sensor network experimental platform show that the proposed algorithm has advantages in tracking accuracy,data transmission and energy consumption.