智能感知与计算研究中心知识库

随着深度神经网络的广泛应用，多目标跟踪这一任务已经取得了巨大的进展。然而，密集场景下的行人跟踪仍然是一个具有挑战性的问题，其主要原因是行人之间严重的遮挡给运动模型和行人表观特征提取带来了巨大的挑战。行人头部跟踪相比于行人全身跟踪遮挡的情况更少，因此行人头部跟踪任务在最近引起了研究者的广泛关注。设计了一种联合身体表观特征的行人头部跟踪模型，该模型总共有两个主要的模块：第一，行人全身检测框内含有更丰富的纹理信息，可以提取更好的表观特征，因此设计了一种动态的由行人头部检测框生成全身检测框的生成网络。第二，行人全身检测框之间存在互相重叠，导致较为严重的遮挡问题。为了能够提取低噪声的行人表观特征，利用人体姿态估计的信息作为引导，使得行人重识别网络更好的关注到非遮挡部分，进而提取到去噪声的行人表观特征。通过上述两个模块，所设计模型在行人头部跟踪的基准数据集上取得了当前最好的效果。此外，该模型还可以应用于行人全身跟踪任务。

Substantial improvement was achieved in the field of multi-object-tracking due to the successful application of deep neural networks. However, tracking pedestrians in crowd remained a challenging problem since great challenge in motion prediction and appearance feature extraction is introduced by severe occlusion. Pedestrian head tracking task was more and more concentrated on because human head part has fewer occlusions than full-body part.  A part-based motion and person Re-ID(Person Re-identification) model which made full use of both the head part and full body for tracking was proposed. Two modules were included in this model. Firstly, a dynamic full-body bounding box generator was designed to make full use of the full-texture feature of the full-body part. Secondly, since severe occlusions exist in full-body bounding boxes, the human pose estimation was used to help appearance feature extraction to focus on the visible part. With the two modules above, the part-based motion and Re-ID model achieves the state-of-the-art result on the pedestrian head tracking benchmark. Besides, this model could also be used for full-body tracking tasks.