基于深度学习的机器人场景识别研究

发布时间：2018-02-04 10:06

本文关键词： 机器人场景识别卷积神经网络深度学习　出处：《沈阳工业大学》2017年硕士论文　论文类型：学位论文

【摘要】：在机器人领域中,机器人到达一个陌生的环境中,如何识别当前环境是计算机视觉领域极其重要的研究问题,也是机器人定位与导航中的一个基本问题,机器人的场景识别研究有助于获取机器人所在工作环境中的实时位姿数据,是机器人实时对当前工作环境地图进行构建的关键步骤。研究者希望机器人能够通过之前的经验自动识别出当前所处环境的类别,这有助于机器人完成接下来的工作任务。传统的场景识别方法的好坏取决于场景识别技术中最关键一步提取场景图像的特征,然而要提取出好的场景图像特征需要花费大量的时间和精力,同时这一过程需要启发式的经验。针对这一问题,本文将深度学习中卷积神经网络模型应用于机器人工作场景识别,它可以自动的从图像中学习到隐藏在其内部的特征数据,这有助于减少人工提取图像特征的工作量。为使机器人具备识别工作场景的能力,本文建立了机器人场景识别系统,并为该系统制定了一个包含多个卷积层、池化层以及全连接层的卷积神经网络结构;系统使用视觉传感器来获取并处理训练卷积神经网络结构时所需的机器人工作场景图像数据集;在获取机器人工作场景图像数据集的过程中使用激光传感器采集环境中的障碍物的距离数据,并结合机器人的位姿系统数据完成机器人在工作场景中的自定位与环境地图的创建工作;另外本文通过多组实验来确定卷积神经网络模型中使用的激活函数类型、池化层的采样方法以及训练时采用的学习率下降方式等,并根据实验最终结果确定了本文网络结构模型采用ReLu激活函数、池化层使用最大池化方法、学习率采用每代都减小的方法以及前两层全连接层都带有Dropout技术。为验证本文卷积神经网络结构模型的有效性,分别使用自己采集的数据集和网络数据集对其进行训练和测试,并对网络结构模型在不同场景类别及数量时的识别效果进行实验验证。实验结果表明本网络测试识别准确率较高,应用在实际机器人系统时在多个角度拍摄机器人工作场景图像,并依次将多角度图像输入训练好的网络结构模型进行场景识别,能完成机器人的实际需求。
[Abstract]:In the field of robot, the robot reaches a strange environment. How to identify the current environment is an extremely important research problem in the field of computer vision, and it is also a basic problem in robot positioning and navigation. The research of robot scene recognition is helpful to obtain the real-time position and pose data of the robot working environment. It is the key step for the robot to construct the map of the current working environment in real time. Researchers hope that the robot can automatically identify the category of the current environment through the previous experience. The traditional scene recognition method depends on the most critical step of scene recognition technology to extract the feature of scene image. However, it takes a lot of time and effort to extract good scene image features, and the process requires heuristic experience. In this paper, the convolution neural network model in depth learning is applied to the robot working scene recognition, which can automatically learn from the image to hide in the internal feature data. This helps to reduce the workload of manually extracting image features. In order to make the robot have the ability to recognize the working scene, a robot scene recognition system is established, and a multi-convolution layer is developed for the system. The convolution neural network structure of the pool layer and the full junction layer; The system uses visual sensor to obtain and process the image data set of robot working scene when training convolutional neural network structure. Laser sensors are used to collect the distance data of obstacles in the environment in the process of acquiring the image data set of the robot working scene. Combined with the position and pose system data of the robot, the self-positioning of the robot in the work scene and the creation of the environmental map are completed. In addition, this paper determines the type of activation function used in the convolution neural network model, the sampling method of the pool layer and the learning rate decreasing method used in the training through multi-group experiments. According to the final results of the experiment, the ReLu activation function is used in the network structure model, and the maximum pool method is used in the pool layer. The learning rate is reduced from generation to generation and the first two layers with Dropout technology. To verify the effectiveness of the convolution neural network structure model in this paper. The data set and network data set are used to train and test them respectively. And the network structure model in different scene categories and the number of the recognition effect of experimental verification. The experimental results show that the network test recognition accuracy is high. When the robot is applied to the actual robot system, the working scene images of the robot are photographed from multiple angles, and the multi-angle images are input into the trained network structure model in turn to recognize the scene, which can fulfill the actual requirements of the robot.
【学位授予单位】：沈阳工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【相似文献】