Researches On Sensor Scheduling And Information Fusion Of Target Tracking In Wireless Sensor Network

Advances in embedded systems, information processing, wireless communication have led to rapid development and wide application of wireless sensor network(WSN). WSN has realized the interconnection of the physical world and the information world leading by Internet, which allows people to get interested information anytime and anywhere. Based on the characteristics of low cost, high redundancy and strong network expansibility, target tracking in WSN has become a widely research in both military and civil filed and therefore has achieved rapid development. This thesis mainly focuses on sensor scheduling technlology and information fusion technology in WSN target tracking. To improve the accuracy and reduce energy consumption of target tracking in WSN, some theoretical and practical methods are put forward. The main work and contributions are as follows:In order to improve the performance of tracking system, it is necessary to determine which sensors should and how to take part in the tracking process because the contribution to estimation of target’s state varys when the different sensors are adopted for tracking task and when the same sensor acts for the cluster head and the ordinary task node. A method is introduced to schedule sensors based on posterior Cramer-rao lower bound(PCRLB) in the tracking scenario that the raw measurement of cluster head and the quantized measurements of ordinary task sensors are used in the process of information fusion. The current cluster head accordingly calculates PCRLB for every combination of a certain number of candidate sensors those are virtually prepared for task nodes next time, which provides a guideline to decide the cluster head and the ordinary task sensors at the next time. The corresponding particle filter algorithm is designed to estimate the target’s state. The experimental results show that the proposed algorithm outperforms the Kullback-Leibler(KL) divergence-based sensor scheduling and random selection of cluster head and ordinary task sensors.Target localization and tracking error mainly comes from multipath and non-line-of-sight transmission in non-ideal communication channel environment. Such environement introduces significant bias in measurements and thus causes large error in traditional positioning techniques. The proposed method embeds the framework of machine learning in target tracking and no longer regards received signal strength indicator(RSSI) as the metric of distance, but uses least square-support vector machine(LS-SVM) to establish the relationship of target’s coordinate and RSSI sequence directly. In the period of target tracking, the model learned leads to a first position estimate of the target under the sensors’ RSSI sequence. The Kalman filter is used afterward to combine the prediction of the target with the first estimate. Simulation results show that the proposed method based on LS-SVM performs better compared with the traditional method based on maximum likelihood estimate(MLE). Besides, smoothed by Kalman filter, tracking accuracy gets significant improvement. However, in the real tracking scenario, dut to the limitation of network energy consumption, tracking delay and communication bandwidth, the data received by the fusion center is usually not raw or complete, but quantized. The error caused by direct quantization is significant when the measurements range widely. Therefore, a new quantization scheme is presented to reduce the bad impact. Firstly, data preprocessing is adopted to shrink range of data. Secondly, the quantization bit number of each sensor’s measurement is decided by its effect on state estimation to improve the information gain of data. The experimental results show the performance of the proposed quantization strategy is superior to the uniform quantization and can be close to the accuracy of using raw measurements in bad channel environements.