基于多变量动态建模的粒子滤波行人跟踪算法实现

Particle filter algorithm for tracking person tracking based on dynamic modeling

对于行人运动模型是线性系统,噪声符合高斯分布,采用边检测边跟踪的卡尔曼滤波算法,试验达到了预期的效果。但在实际中行人的随机行走具有很大的不确定性,不一定是线性系统和高斯分布,此时利用Kalman滤波就会导致跟踪失败。研究了基于先检测后跟踪的加权颜色直方图为匹配模板,基于动态建模的粒子滤波实现对行人的有效跟踪。在初始帧利用AdaBoost算法确定行人的位置、大小等状态信息,以行人矩形框内的加权颜色直方图作为跟踪的目标模板,初始化粒子集。在后续的视频图像中,利用粒子滤波算法实现行人跟踪。结果表明,即使在目标有遮挡、阴影等复杂噪声背景下,提出的方法也能很好地跟踪到视频序列中行人。

We study on particle filtering algorithm for tracking person based on the weighted color histogram.In the thesis,we use the method with detecting person firstly and tracking it secondly.First,we apply the pedestrian detection algorithm to rectangle the persons in the images and compute the status information,like position,scale,and so on.Then we compute the weighted color histogram in the rectangle region as target template.Then we initialize the particle set according to the start status.The random work model is used as state equation in the initial tracking stage to propagate the state of particles.Then we compute the weighted color histogram of the each particle as observation data,and also compute the similarity coefficient between them and the target template.According to the similarity coefficient,each particle is weighted by the new value.Then all weighted particles are summed to get the final state estimation to be output.Finally,the particles were re-sampled to reduce the degradation.Since the persons walk in the uncertainty and random,one state equation is not effective to be used to track the target during the whole procedure.So we propose the method with dynamical state equation update based on time series to reduce the estimation error.The experimental results show that the proposed method can well track the video sequences in the presence of complex background such as occlusion and shadow.