基于视觉与Wi-Fi双层特征地图的智能手机室内定位方法

Smartphone indoor positioning method based on vision and Wi-Fi double-layer feature map

关键词:  室内定位  智能手机  双层特征地图  隐马尔可夫模型  Wi-Fi定位  视觉定位

Key words:  indoor positioning  smartphone  double-layer feature map  HMM  Wi-Fi positioning  visual positioning

摘要:

为解决室内环境中Wi-Fi定位精度低、视觉定位稳定性差等问题,提出了一种融合视觉与Wi-Fi的双层特征地图模型（vision-CSI map,V-CSI map),并提出基于隐马尔可夫模型（Hidden Markov model,HMM）的智能手机定位方法。V-CSI地图模型既包含基于栅格的信道状态信息（channel state information,CSI）指纹特征,同时也包含以稀疏安全出口为路标的视觉特征,并通过参考位置关联完成地图构建。本文将基于V-CSI特征地图的定位问题转化为HMM问题。首先完成安全出口标志检测与视觉特征匹配,实现视觉定位,将定位结果对HMM状态进行初始化；接着,利用CSI指纹匹配完成发射概率建模,通过高斯模型完成基于运动约束的状态转移概率建模；最后,通过前向算法求解HMM最优匹配状态,进而计算位置。在6 000平方米办公楼和3 600平方米地下停车场对本文算法与模型进行验证。实验结果表明,两种典型室内场景下,本文算法平均定位误差约为1.0 m,单次定位时间约170 ms；相比于单一CSI定位,平均定位误差减少56%以上,说明该算法能够有效提升室内定位的准确性与鲁棒性。

Abstract:

We proposed a smartphone positioning method by formulating the positioning problem as an HMM(hidden markov model,HMM) based on the proposed double-layer feature map consisting of visual and Wi-Fi features(vision-CSI map,V-CSI map) to solve the issue of low accuracy and poor stability in indoor environment. The V-CSI map is modeled by encoding CSI fingerprint features based on grid and visual features of sparse safety exits as well as association locations. The location problem based on the V-CSI feature map is solved as HMM problem in the method. First, the safety exit sign detection and visual feature matching are completed in the visual positioning phase, and the positioning results are employed to initialize and reinitialize the states of HMM. Subsequently, CSI fingerprint features are matched with that of the V-CSI map to complete the emission probability, and the state transition probability is computed by modeling motion constraint with Gaussian model. Finally, the optimal state is derived from the forward algorithm, and the position of the smartphone is readily determined from the weighted average of the closest states. In the experiment, the proposed method is verified in an office building of 6 000 square meters and an underground parking lot of 3 600 square meters respectively. Experimental results show that the average positioning error of the algorithm is about 1.0 m, and the time of a single positioning is about 170 ms in the two typical indoor scenes. Compared with only CSI positioning methods, the average positioning error of our proposed method is reduced by more than 56%. The outstanding performance of experimental results also illustrates that our proposed method can improve the accuracy and robustness of indoor positioning.