一种基于深度学习目标检测的长时目标跟踪算法

PDF下载HTML

Title:

A long-term object tracking algorithm based on deep learning and object detection

作者:

邵江南^1,2, 葛洪伟^1,2

1. 江南大学 江苏省模式识别与计算智能工程实验室，江苏 无锡 214122;
2. 江南大学 物联网工程学院，江苏 无锡 214122

Author(s):

SHAO Jiangnan^1,2, GE Hongwei^1,2

1. Jiangsu Provincial Engineering Laboratory of Pattern Recognition and Computational Intelligence, Jiangnan University, Wuxi 214122, China;
2. School of Internet of Things, Jiangnan University, Wuxi 214122, China

关键词:

目标跟踪; 长时跟踪; 神经网络; 卷积特征; 类不均衡问题; 损失函数; 特征提取; 深度学习

Keywords:

object tracking; long-term tracking; neural network; convolutional features; class imbalance problem; loss function; feature extraction; deep learning

分类号:

TP391.4

DOI:

10.11992/tis.201910029

摘要:

针对长时目标跟踪所面临的目标被遮挡、出视野等常常会导致跟踪漂移或丢失的问题，基于MDNet提出一种深度长时目标跟踪算法(long-term object tracking based on MDNet, LT-MDNet)。首先，引入了一种改进的收缩损失函数，以解决模型训练时正负样本不均衡的问题；其次，设计了一种高置信度保留样本池，对在线跟踪时的每一帧的有效并且置信度最高结果进行保留，并在池满时替换最低置信度的保留样本；最后，在模型检测到跟踪失败或连续跟踪帧数达到特定阈值时，利用保留样本池进行在线训练更新模型，从而使模型在应对长时跟踪时保持鲁棒和高效。实验结果表明，LT-MDNet在跟踪精度和成功率上都展现了极强的竞争力，并且在目标被遮挡、出视野等情况下保持了优越的跟踪性能和可靠性。

Abstract:

Aiming at the problem of tracking drift or loss caused by the occlusion and the out-of-view of the target in long-term tracking, this paper proposes a new deep, long-term tracking algorithm based on MDNet (LT-MDNet). First, an improved shrinkage loss function is introduced to solve the problem of the positive-negative class imbalance in the model training. Second, a high confidence retention sample pool is designed to retain the valid and highest confidence results of each frame during online tracking and to replace the lowest confidence retention samples when the pool is full. Finally, when the model detects a tracking failure or when the continuous tracking frame number reaches a specific threshold, the reserved sample pool is used for online training to update the model to make the model robust and efficient in dealing with long-term tracking. Experimental results show that LT-MDNet is highly competitive in its tracking accuracy and success rate and maintains excellent tracking performance and reliability in the case of target occlusion and out-of-view.