复杂场景下基于局部运动补偿的目标检测  

Object detection based on local motion compensation in complex scenes

作　　者：李响[1] 谭南林[1] 王天雷[1] 苏树强[1] 

机构地区：[1]北京交通大学机械与电子控制工程学院,北京100044

出　　处：《仪器仪表学报》2014年第7期1555-1563,共9页Chinese Journal of Scientific Instrument

基　　金：中国铁路总公司科技研究开发计划(2012x0007-c)资助项目

摘　　要：为了应对视频图像中的复杂场景,提出一种基于局部运动补偿的二次分割目标检测方法。首先,针对光照、阴影变化及目标伪装色的影响,提取了视频图像的像素级色度特征信息和局部纹理梯度特征。同时建立多维高斯分布的背景概率模型以及均匀分布与高斯分布相结合的前景概率模型,并根据贝叶斯决策理论对当前帧图像进行预分割得到静态背景和动态像素。其次,为了从动态像素中进一步区分出实际的前景目标,采用块匹配运动估计算法得到各动态像素点的帧间运动矢量,并将其补偿回运动初始区域。随后在动态像素的特征向量中引入运动矢量特征,建立局部运动补偿后的时空域联合前、背景概率模型,进行二次分割得到实际前景目标。最后给出了模型的初始化过程,并提出一种各模型参数的自适应实时更新策略以提高算法的鲁棒性。实验结果表明,该方法能够良好地适应不同的复杂场景,且具有较好的检测效果和实时性能。A twice segmentation method for object detection based on local motion compensation is proposed to deal with complex scenes in video sequences. Firstly, aiming at the influence of the illumination- shadow changes and object camouflage, the pixel- level chrominance features and local texture gradient features of video images are extracted. Meanwhile, the background probabilistic model with multivariate Gaussian distribution and the foreground probabilistic model with multivariate mixture of uniform distribution and Gaussian distribution are built. According to Bayesian decision theory, the current frame image is then pre - segmented to obtain static background pixels and moving pixels. Secondly, in order to further distinguish the real foreground objects from the moving pixels, a block matching motion estimation algorithm is adopted to calculate the inter - frame motion vectors of various moving pixels and compensate them back to the starting area. So the motion vector can be fused into the moving pixel feature vector, and the spatio - temporal background and fore- ground probabilistic models are established based on local motion compensation, the second segmentation is performed to get the final foreground object. At last, the model initialization procedure is described, and a real - time adaptive model parameter update strategy is proposed to improve the robustness of the algorithm. The experiment results show that this method is more suitable for the object detection in different complex scenes, and can achieve better detection result and real -time performance.

关 键 词：目标检测 复杂场景 局部运动补偿 二次分割 时空域联合建模 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TH7[自动化与计算机技术—计算机科学与技术]