随着深度学习在计算机视觉任务中的广泛应用，在多种视觉任务，例如图像分类和人脸识别，深度学习的算法不仅准确率超过了人类水平，而且在实际应用场景中取得了巨大的成功。 现阶段的深度学习主要是基于分类的预测学习，特征表示能力强，但决策能力弱，不能高效地处理协同感知与决策任务。随着DeepMind公司将强化学习和深度学习相结合，在视觉分析的协同感知与决策上取得了巨大的进展。由此，本文将深度强化学习引入到视觉目标跟踪与再识别任务中，协同解决视觉特征表示与视觉任务决策。本文主要研究工作和创新点具体如下：针对单目标跟踪中目标高效定位和状态决策的问题，提出了一种基于深度强化学习的连续迭代位移的单目标跟踪方法，并利用决策网络采取相关动作来更新目标状态，而不是通过大量的候选框采样然后进行在线分类来实现目标的定位，该方法鲁棒性强计算效率高，可以处理大变形和快速运动的情形。实验结果表明，所提出的方法鲁棒性强计算效率高。针对多目标跟踪系统对检测结果依赖性高、决策能力弱的问题，提出了一种基于多智能体强化学习的多目标跟踪学习方法。所提出的方法将每个目标视为一个智能体，并通过预测网络对其进行轨迹预测，通过决策网络利用不同智能体和环境的协作交互来寻求整体的最佳跟踪结果。实验结果表明，所提出的方法可以有效提升多目标跟踪系统的准确率。针对摄像机网络下行人再识别问题，提出了一种用于相机网络中行人再识别的一致性保持特征学习方法。利用深度强化学习技术在学习特征表示的同时利用一致性约束信息，来获得整个摄像机网络中行人的最优匹配，并且对表示网络进行端到端的训练。实验结果表明，所提出的方法获得了显著的性能改进，并且大大超过了现有技术。针对车辆再识别中判别特征挖掘问题，提出了一种通过深度强化学习对车辆进行再识别的注意感知判别特征挖掘方法。 所提出的方法通过多分支注意力模块来定位车辆的高判别区域，并通过策略网络来为各个车辆自适应地选择可用于识别的判别区域。 在公开数据集上的实验结果表明，所提出的方法达到了最先进的性能。

With the wide application of deep learning in computer vision tasks, in a variety of vision tasks, such as image classification and face recognition, deep learning algorithms not only exceed the human level in accuracy, but also have achieved great success in practical application scenarios. . At this stage, deep learning is mainly based on classification-based predictive learning. It has strong feature representation but weak decision-making ability and cannot efficiently handle collaborative perception and decision-making tasks. Collaboration, here will be in-depth reinforcement learning dating between visual target tracking and re-recognition tasks, and collaboratively solve visual feature representation and visual task decision-making. The main research work and innovations here are as follows:Aiming at the problem of efficient target positioning and state decision-making in single target tracking, a single target tracking method based on continuous iterative displacement based on deep reinforcement learning is proposed, and the decision network is used to take relevant actions to update the target state, rather than through a large number of The candidate frame of the sample is sampled and then classified online to achieve target positioning. This method is robust and has high computational efficiency, and can handle large deformations and fast motion situations. Experimental results show that the proposed method has strong robustness and high computational efficiency.Aiming at the problem of high dependence on detection results and weak decision-making ability of the multi-target tracking system, a multi-target tracking learning method based on multi-agent reinforcement learning is proposed. The proposed method regards each target as an agent, and predicts its trajectory through the prediction network, and seeks the overall best tracking result through the collaborative interaction of different agents and the environment through the decision-making network. Experimental results show that the proposed method can effectively improve the accuracy of the multi-target tracking system.Aiming at the problem of pedestrian re-identification in camera networks, a consistent and consistent deep learning method for pedestrian re-identification in camera networks is proposed. Using deep reinforcement learning technology to learn feature representation while using consistency constraint information to obtain the optimal matching of pedestrians in the entire camera network, and to perform end-to-end training on the representation network. Experimental results show that the proposed method has achieved significant performance improvement and greatly exceeds the existing technology.Aiming at the problem of discriminative feature mining in vehicle re-identification, a method of attention perception discriminant feature mining that re-identifies vehicles through deep reinforcement learning is proposed. The proposed method uses a multi-branch attention module to locate the high discriminative area of ?​the vehicle, and uses a strategy network to adaptively select the discriminant area that can be used for identification for each vehicle.