激光与光电子学进展, 2019, 56 (12): 121501, 网络出版: 2019-06-13
基于核相关滤波的多目标追踪 下载: 1391次
Multiple Object Tracking Based on Kernelized Correlation Filter
机器视觉 多目标跟踪 核相关滤波 交并比 machine vision multiple object tracking kernelized correlation filter intersection-over-union
摘要
为了克服核相关滤波(KCF)只根据目标外观模型追踪时准确性低的不足,融入运动模型,计算了检测目标框和预测目标框的交并比(IOU)。通过匈牙利算法,确定了目标间的最优关联。KCF和IOU模型都具有快速响应的特点,因此算法可满足在线处理数据的要求。在公开的2DMOT2015、MOT16数据集上进行实验,将所提方法与其他优秀方法相比,在保证30 frame/s以上处理速度的同时,追踪准确性提高10%以上。
Abstract
A motion model is used to overcome the shortcoming of the low tracking accuracy of a kernelized correlation filter (KCF) based only on a target appearance model. The intersection-over-union (IOU) between the detection target bounding box and the predicted target bounding box was calculated. The optimal correlation among the targets was determined using the Hungarian algorithm. Both the KCF and IOU models are characterized by fast responses; therefore, the algorithm has the ability to process data online. The experiments were conducted on the public 2DMOT2015 and MOT16 datasets. Compared with the other state-of-the-art method, the tracking accuracy of the proposed method is higher than 10% while ensuring a processing speed of 30 frame/s or faster.
刘欢, 李春庚, 安居白, 魏帼, 任俊丽. 基于核相关滤波的多目标追踪[J]. 激光与光电子学进展, 2019, 56(12): 121501. Huan Liu, Chungeng Li, Jubai An, Guo Wei, Junli Ren. Multiple Object Tracking Based on Kernelized Correlation Filter[J]. Laser & Optoelectronics Progress, 2019, 56(12): 121501.