基于多特征融合的多摄像机人体跟踪方法

在非重叠视野摄像机网络中, 因视觉盲区等因素的存在, 难以对人体目标进行准确可靠的持续跟踪, 为此, 提出一种融合主颜色特征、 纹理特征和时空拓扑特征的目标跟踪算法。该算法将人体区域分割成上、 中、 下3个目标子块, 分别利用最近邻聚类算法提取每个目标子块的主颜色信息, 并计算主颜色匹配率； 通过提取目标的空间纹理特征获得纹理匹配率； 最后通过融合计算人体外观匹配模型。同时, 根据目标关联信息的累计统计信息, 采用增量学习思路建立和更新摄像机网络的时空拓扑关系。实际场景的实验表明, 该算法能有效地对非重叠视野多摄像机网络中出现的人体目标进行连续跟踪, 并随系统的持续运行和监控区域中新目标的不断出现, 其跟踪准确度也随之提高。

Human Tracking Method Based on Multiple Features Fusion Across Multiple Cameras

In the camera network with non-overlapping FOVs （ Field of Views）, due to the factors such as the visual blind spot, it is difficult to track human continuously across multiple cameras. A human tracking method fusing the main color feature, textual feature and spatio-temporal topology feature is proposed. A SNNC （Sorted Nearest Neighbor Clustering） algorithm is adopted to extract the main color feature from the three human body parts which is head part, torso part, and legs part, and the matching rate is acquired. The spatial textual feature of the human are extracted to obtain the textural similarity. Combined with the two features above, the human appearance matching mode is constructed. Based on the statistic object correspondence information, the incremental learning method is exploited to construct and update the spatio-temporal information. The experiments prove that the proposed human tracking method can track the objects continually in camera network with non-overlapping FOVs. And the accuracy become higher over time as new observations are accumulated without supervised input.