综合信息系统研究中心知识库

跟随技术是移动机器人重要的研究方向之一，在各种跟随技术的解决方案中，基于视觉传感器的方法凭借其精度高、适应性强、价格便宜、便于携带等特点成为了首选，其主要通过视觉跟踪算法来实时获取目标位置信息，进而可以获取目标的距离、角度等信息，实现移动机器人的跟随功能，所以提升视觉目标跟踪算法的精度和鲁棒性对于移动机器人稳定跟随有着重要意义。近些年来，基于相关滤波的视觉目标跟踪算法由于其精度高、求解高效等特点成为了视觉跟踪领域中的主流算法，但是在实际应用中，由于需要面对外观改变、局部遮挡、相似目标干扰、背景干扰等场景，相关滤波跟踪算法的性能还有待提高。本文针对移动机器人在实际场景可能遇到的一些有挑战性的场景和相关滤波跟踪算法本身存在的一些缺陷，以提升算法的精度和鲁棒性为目标，从跟踪模型设计和特征选择两个方面对相关滤波跟踪算法进行了改进，并将算法应用到移动机器人平台上来验证其在实际场景中的性能。主要工作总结如下：

        1. 本文从正则项和上下文学习出发，提出了一种基于时空正则项和自适应感知上下文权重的相关滤波跟踪算法。该算法提出的自适应感知上下文权重的方法使滤波器可以利用上下文样本进行训练，同时这些样本的权重在训练过程中可以自动地调节，可以有效地应对相似目标干扰场景。同时，该算法提出的基于中心注意力机制的空间正则项以及自动调节时间正则项参数的方法可以使跟踪器更好地缓解边界效应问题以及应对遮挡带来的模型退化问题。

        2. 本文从特征选择的角度出发，提出了一种基于空间特征自适应选择的上下文感知相关滤波跟踪算法。该算法提出的L1正则项可以在空间上挑选出判别性强的特征，提升滤波器的判别性能，可以很好地解决背景扰、外观改变等问题。同时，加入的时间约束项可以使滤波器在时间上具有很好的平滑性，可以很好地处理模型退化等问题。

        3. 本文针对移动机器人平台设计了一个行人跟随系统来验证两种相关滤波跟踪算法在实际应用中的性能。整个系统利用深度相机获取信息，采用两种相关滤波跟踪算法实时获取行人位置，然后通过计算可以获得行人相对于机器人的角度信息和距离信息，最后用PID控制器校正机器人的线速度和角速度，实现跟随行人的功能。

        本文提出的两种基于相关滤波的目标跟踪算法在公开数据集和移动机器人平台上进行了实验，实验结果证明了其在具有挑战性的实际场景下具有良好的鲁棒性，为解决移动机器人视觉跟踪问题提供了有益的思路和方法。

Following technology is one of the important research directions of mobile robots. Among various following technology solutions, the method based on vision sensors has become the first choice due to its high accuracy, strong adaptability, low price, and easy portability. The visual tracking algorithm can obtain the target position information in real time, and then can obtain the target's distance, angle and other information to realize the following function of the mobile robot. Therefore, improving the accuracy and robustness of the visual object tracking algorithm is of great significance for the stable following of mobile robots. In recent years, the visual object tracking algorithm based on correlation filter has become the mainstream algorithm in the field of visual tracking due to its high accuracy and efficient solution. However, in practical applications, because of facing various complex scenes, such as appearance change, partial occlusion, similar target interference, background interference, the performance of tracking algorithms based on correlation filter still needs to be improved. This article focuses on some challenging scenes that mobile robots may encounter in actual scenes and some defects in the tracking algorithm based on correlation filter itself, with the goal of improving the accuracy and robustness of the algorithm, we improved the tracking algorithm based on correlation filter itself, with the goal of improving the accuracy and robustness of the algorithm from the two aspects of tracking model design and feature selection. And the algorithms are applied to the mobile robot platform to verify its performance in actual scenes. The main work is summarized as follows:

        1. From the perspective of regular term and context learning, we propose a correlation filter tracking algorithm based on temporal and spatial regular terms and adaptive context-weight-aware method. The adaptive context weighting method proposed by this algorithm enables the filter to use context samples for training. At the same time, the weights of these samples can be automatically adjusted during the training process, which can effectively deal with similar target interference scenarios. At the same time, the spatial regularization based on the central attention mechanism and the method of automatically adjusting the temporal regularization parameters proposed by the algorithm can make the tracker better alleviate the boundary effect problem and deal with the model degradation problem caused by occlusion.

        2. From the perspective of feature selection, we propose a context-aware correlation filter based tracking algorithm via adaptive selection of spatial features. The L1 regular term proposed by the algorithm can select highly discriminative features in space, improve the discriminative performance of the filter, and can well solve the problems of background interference and appearance change. At the same time, the added time constraint items can make the filter have good smoothness in time, and can deal with problems such as model degradation.

        3. A pedestrian following system is designed for a mobile robot platform to verify the performance of two related filtering tracking algorithms in practical applications. The entire system uses the depth camera to obtain information, uses two correlation filter tracking algorithms to obtain the pedestrian position in real time, and then calculates the angle and distance information of the pedestrian relative to the robot. Finally, the PID controller is used to correct the robot's linear velocity and angular velocity, and then realize the function of following pedestrians.

        The two visual object tracking algorithms based on correlation filter proposed in this paper have conducted a large number of experiments on public data sets and mobile robot platforms, and the experimental results have proved their effectiveness. It has good robustness in challenging actual scenarios, and provides useful ideas and methods for solving the problem of mobile robot visual tracking.