Research On The RSS Localization Accuracy Enhancement In Complex NLOS Environments

None-line-of-sight(NLOS) biases are common in localization applications, which degrades the location estimation accuracy seriously. The troublesome of NLOS biases exists in any localization system based on measurements such as TOA(time-of-arrival), AOA(angle-of-arrival), or RSS(received-signal-strength) etc. Although the RSS-based location error is at meters, the RSS-based localization system is still preferred in some applications without critical accuracy constraints, such as vehicle localization in parking stations etc, due to the convenience and lowest hardware cost of obtaining RSS measurements. Therefore, it is important to research on the issue about how to improve the localization accuracy in complex NLOS environments based on RSS measurements. The identification and mitigation of NLOS biases are challenging due to the easy fluctuation of radio signal and the complex statistical characteristics of RSS measurements.NLOS biases usually vary dynamically in complex NLOS environments, which makes the offline machine learning method of localization be likely in vain(in spite of the huge effort of obtaining the offline training set). On the contrary, the localization method based on the log-path-loss propagation model need not any offline learning effort, thus more convenience for rapid implementation. The propagation-model-based localization method is adopted in this thesis, and to be researched on how to improve the localization accuracy in complex NLOS environments, with only the prior knowledge of RSS measurements and the coordination of anchors.Firstly, the approximate statistical distribution of NLOS biases should be obtained. The real distribution of NLOS biases are completely unknown and variant in complex NLOS environments, and the exact distribution for NLOS biases is usually hard to obtain. Although the precision distribution model can be obtained by some methods like KDE(Kernal Distribution Estimation) etc, with offline learning effort, the model may likely to fail in online applications due to the dynamic characteristics of NLOS biases. Therefore, the classic statistical distribution is adopted to approximate the real distribution of NLOS biases. In experiments, the normal distribution, the uniform distribution, and the exponent distribution are used to model the NLOS biases respectively, then the maximum likelihood estimator(MLE) is formulated accordingly, finally location estimations are performed based on the according MLE. Experimental results reveal that the normal distribution is the most suitable choice.Secondly, the negative effect of GDOP(Geometric dilution of precision) should be minimized. A better localization performance would be achieved in a network with good GDOP as for the same localization algorithm. Therefore how to obtain a good GDOP when selecting optimal locations to deploy anchors is an important issue. The issue is modeled to be a QP(quadratic programming) problem in this thesis, which can be solved effectively. Thus the near optimal geometric deployment of anchors can be obtained conveniently, and served for better localization performance later.Thirdly, the specific localization algorithms are developed. There is a dilemma in the propagation-model-based localization. Generally, the path loss exponent of the propagation model is unknown and varying in complex NLOS environments. However, on the other hand, the path loss exponent value is needed to perform location estimation. Although it is feasible to estimate the path loss exponent jointly with the unknown target coordination, the estimation accuracy will decrease due to the increasement of unknown variables. In order to avoid this dilemma, the relationship between the distance vector and the RSS measurements vector are exploited for obtaining the coarse location estimation of the target, then the approximate value of the path loss exponent is obtained based on the coarse location estimation, thus the final location estimation can be performed.In order to eliminate the effect of NLOS biases and further improve the localization accuracy, the method by using the negative exponent function to map the loclization residuals into the weight of anchors is proposed, based on the statistical correlation between the localization reisuals and the localization errors. The bigger is the weight, the less influence on the estimation errors caused by the NLOS biases. On the contrary, the smaller is the weight, the more impact on the estimation error. Therefore, the weighting strategy can effectively lower the influence caused by the NLOS biases. However, this method still can not eliminate the NLOS biases completely, which is a strictly difficult problem for no prior knowledge of NLOS biases and only prior knowledge of RSS measurements and the anchors’ coordinations.Finally, the path loss exponent of the propagation model is modified into a function of the distance, which is inspired from the path loss characteristics in the Zigbee-based localization systems. The localization accuracy can be enhanced by the MLE based on the modified propagation model, especially when the target is close to the anchors.