基于深度学习的多目标跟踪算法研究

[摘要] 提出了一种基于深度学习的多目标跟踪算法。首先，通过GoogLeNet+长短期记忆网络（LSTM）模型进行目标检测，以获得准确的目标检测结果；其次，直接根据目标检测的特征图对检测目标进行深度特征的提取，深度特征相比于传统特征可以更准确地反映检测目标的外观特征，因此可以有效提高跟踪的准确性。此外，还在传统数据驱动马尔科夫蒙特卡洛（DDMCMC）算法的基础上，提出了层次的数据驱动马尔科夫蒙特卡洛（HDDMCMC）算法，可以进一步提高多目标跟踪的准确性。实验结果证明了所提出算法的有效性。

[关键词] 多目标跟踪；深度学习；目标检测；MCMC算法

[Abstract] In this paper, a multi-target tracking algorithm based on deep learning is proposed. Firstly, GoogLeNet + long short-term memory (LSTM) model is used to obtain accurate object detection results. Secondly, the feature map of object detection is directly used to extract the deep feature for tracking. Compared with the traditional feature, the deep feature can reflect the appearance of objects more accurately, which could improve the tracking accuracy effectively. What's more, based on the traditional Data Driven Markov Chain Monte Carlo (DDMCMC) algorithm, the Hierarchical Data Driven Markov Chain Monte Carlo (HDDMCMC) algorithm is proposed to further improve the tracking accuracy. The experiment results prove the effectiveness of our algorithm.

[Keywords] multiple object tracking; deep learning; object detection; MCMC algorithm