应用自适应多测量融合粒子滤波的视场跟踪

针对矿井跟踪视场中由于单一线索对目标特征描述缺乏可分性以及多线索融合策略对场景变化缺乏自适应性导致人员跟踪失效的问题,提出了基于自适应多测量融合粒子滤波的矿井人员跟踪算法。将粒子邻域光流统计信息表征的运动性作为线索建立运动光流直方图模型,并与颜色相融合建立多观测模型。将单观测估计状态粒子区域与融合估计粒子区域的质心距离作为单观测模型贡献率度量因子,定义了观测权值自适应策略,实现了粒子观测模型与跟踪目标状态特征的同步变化;通过建议重采样函数对粒子落入低观测似然时进行有效的采样补偿,增强了跟踪的鲁棒性。实验结果表明,本算法能够有效地解决矿井跟踪视场下(背景复杂)由于场景变换而导致跟踪目标丢失的问题;将本文算法与基于颜色和基于颜色与帧差分融合的粒子滤波算法做状态估计均方误差比较,结果表明,状态估计准确率提高了1.57倍。

Visual tracking with adaptive multi-cue fusion particle filter

As the target-tracking in coal mines using a single-cue lacks discrimination of target features and strategies using the multi-cue fusion lack the adaptation to changes of scene, a novel particle filter algorithm based on adaptive multi-cue fusion models was proposed for object-tracking.An optical flow histogram was established based on particle motion, then,the optical flow was fused with color information to obtain a multi-cue based observation model. An adaptive strategy of observation model weights was implemented by taking the centroid distance between the single-cue observation model and multi-cue fusion model as the contribution factor of the single-cue observation model. When it was implemented, the particle observation model would change with the object characteristics.The particle re-sampling was achieved by a proposal re-sampling when weights of single-cue observation model were all below a threshold. The results show that the tracking algorithm is an effective solution to tracking failure due to changes of scene in coal mines.The accuracy of estimation has increased by 1.57 times as compared with those of other particle filter algorithms.