面向鲁棒目标跟踪的小样本数据下CNN模型学习方法研究

Recently, discriminative methods based on deep convolutional neural networks (CNNs) have dominated visual tracking. However, the lack of large-scale training data, compared to that of image classification or object detection, greatly limits the trackers' performance. This project studies the training of DCNNs with limited data. On one hand, for the offline training, we (1) learn attribute-based representations which lie in existing training data so as to train effective CNNs without additional data; (2) Transfer the image style of visual tracking data to that of closely related tasks, such as object detection in order to increase data diversity while decreasing their distribution difference and hence alleviate over-fitting. On the other hand, for the online training, we (3) learn to generate the future potential appearance of the target object in order to enhance the model fine-tuning; (4) explore multi-domain multi-task meta-learning to model the learning of learners’ parameters so as to rapidly update learners during test. This project will provide useful technique and theory support for artificial applications including automatic drive, human-computer interaction, smart surveillance, and intelligent traffic control.

近年来，基于深度卷积神经网络的判别式分类模型成为目标跟踪领域的主流方法。然而，相较于图像分类、物体检测等任务，目标跟踪领域训练数据十分匮乏，这极大地限制了深度卷积神经网络的性能。为解决该问题，本项目研究面向目标跟踪的小样本数据下深度卷积神经网络的学习和泛化问题。针对线下训练阶段数据匮乏的问题，本项目（1）研究基于视频属性自动发现的特征表示学习，以增强CNN模型学习能力；（2）研究面向目标跟踪的图像风格迁移模型，在提高数据多样性的同时降低源域与目标域的数据分布差异，降低CNN模型过拟合。针对测试阶段仅有单帧图像可用的问题，本项目（3）研究基于图像生成的目标跟踪方法，实现模型在适配过程中融入跟踪目标未来的潜在表观；（4）研究基于元学习的多域多任务模型，建模参数学习过程，增强模型泛化能力。本项目的研究成果将为自动驾驶、人机交互、智能监控等人工智能应用提供理论依据和技术方案。

长期以来，目标跟踪是计算机视觉研究的重点，热点与难点。本项目面向深度学习环境下训练样本匮乏的关键问题，以样本有效性、帧内特征表示、帧间时空特征以及数据增广等为切入点，在卷积神经网络框架下，结合统计学习、对抗学习等理论与方法，完成以下目标：建立并完善基于视频统计特性的有效样本挖掘；有效利用视频帧内空间特征，兼顾时间维度动作延续性和完整性，增强跟踪目标的特征表示；基于自监督学习实现非对称样本下可控图像编辑与增广。本项目预期降低深度学习对训练样本数量的过度依赖，对促进目标跟踪理论与方法的实际应用意义重大。