因式分解卷积运算的多尺度目标跟踪算法

Multiscale target tracking based on factorization convolution

针对因受到尺度变化、光照变化、形状变化以及相似目标等因素的干扰,目标跟踪过程出现漂移或算法过拟合现象等问题,提出了一种基于因式分解卷积运算的多尺度卷积运算.采用含有类似anchors机制的深度检测模型SSD,提取不同宽高比尺寸的特征减少漂移情况的出现,利用紧凑的样本集生成模型和优秀的更新策略等优点有效地解决了过拟合问题.结合ECO算法中的因式分解法提高在光照变化、尺度变化、遮挡以及背景杂波等方面跟踪效果.实验结果表明:该目标跟踪算法具有较强的鲁棒性和较高的跟踪成功率.研究结论可以提高目标跟踪算法的精确性、实时性.

Aiming at the problem that the target tracking algorithm often drifts in the process of tracking under the interference of various interference factors,or the phenomenon of over-fitting because of the large model.This paper proposes a multi-scale target tracking method based on factorization convolution operation.In order to reduce the drift effectively,this method extracts the features of different width-height ratio sizes through the depth detection model SSD with anchors-like mechanism,and combines the ECO which belongs to the discriminant model.The results show that our tracking method has better tracking accuracy and robustness under illumination variation,scale variation,occlusion and background clutters.Its compact sample set generation model and excellent update strategy effectively solve the over-fitting.