面向智能车的地下停车场环视特征地图构建与定位

Visual map from around view system for intelligent vehicle localization in underground parking lots

针对地下停车场环境GPS信号缺失的问题,本文在环视特征地图构建的基础上,提出基于二阶马尔科夫模型的粒子滤波定位算法(Markov model-particle filter,MM-PF),实现智能车在地下停车场环境中的高精度定位。在该模型中,环视特征地图节点被定义为粒子,查询图像被定义为观测数据。在状态转移过程中,引入二阶马尔可夫模型,对短时间车辆运动进行建模,构建状态转移模型。利用图像的全局特征建立当前车辆获取的图像与各粒子(环视地图节点)之间的匹配关系,从获取的汉明距离建立粒子权重分布模型,可以大幅提高系统的计算效率。当前车辆的位置由局部特征匹配获得。选取两个典型的地下停车场场景对本文算法进行验证,在选取的两个场景中,本文算法平均定位精度小于0.38 m,定位误差均方差小于0.29 m,定位误差在1 m以下的概率不低于95.4%。试验结果表明:本文所提出的二阶MM-PF算法能够将车辆的运动信息与视觉信息相融合,相较于对比算法,定位精度与稳健性得到大幅提高。

In view of the lack of GPS signal in the underground parking lots,the second-order Markov model and particle filter(MM-PF)method for intelligent vehicle localization in underground parking lots is proposed based on the construction of a visual feature map from around view.In the method,the nodes of the visual feature map are defined as particles,while the query images are defined as observation data.In the process of state transition,the second-order Markov model is introduced to model the motion of the vehicles in a short time.In addition,holistic features are employed to establish the matching relationship between the query image and each particle(visual feature map node)by assigning particle weights based on Hamming distance.In the experiments,two typical underground parking lots are selected to verify the method.In both scenarios,the mean error of localization is less than 0.38 m,the mean square error is less than 0.29 m.The probability of positioning error below 1 m is not less than 95.4%.Experimental results demonstrate that the proposed method can integrate both the motion and visual features to enhance localization performance.Experimental results also show that the proposed method outperforms state-of-the-art ones in terms of localization accuracy and robustness.