基于卷积神经网络的移动机器人室内物体检测

发布时间：2018-04-24 18:17

本文选题：计算机视觉 + 深度学习　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：视觉识别技术是机器人能够与人类协同工作的的核心技术之一。过去的计算机视觉系统,在有约束的环境中表现良好,但不适用于复杂多变的工业环境,因此很多任务需要训练有素的工人来完成。近年来,基于机器学习的计算机视觉技术,极大的提升了现实世界中的自然图像分类的识别率。但是这些方法依赖大量的训练数据(数以百万计的训练图像),需要很长的训练时间(几小时到几天),使这些机器学习算法不适用于实时的通用机器人。本课题使用卷积神经网络进行物体识别任务(识别给定图像中的物体)。为了避免长时间的复杂训练,我们将捕获的图像输入预训练好的卷积神经网络中,并将卷积神经网络的输出作为表征图像中目标的特征向量。首先利用摄像机对目标进行跟踪,再将特征向量序列存储在内存中,最后用支持向量机对特征进行训练,从而实现对目标的分类。支持向量机可以在较短的训练时间内进行训练,从而满足实时性要求。本文测试了上述视觉系统在现实世界室内环境中的识别率。利用摄像机和运动检测,设计的系统可以跟踪在视野中显示的物体。如果用户启动了学习模式,系统将为这个新对象训练一个新的支持向量机。如果用户启动了预测模式,系统会基于已经训练过的模型猜测对象的身份。本文也研究了仅使用视觉反馈对对象进行分类的可能。系统会检查给定对象的特征向量与任何已知的目标特征向量的特征向量之间相似性,相似性通过测量两者之间的欧氏距离得到。最后,我们使用标准数据库测试提出的算法,并与其它数据库对比。
[Abstract]:Visual recognition is one of the core technologies in which robots can work together with human beings. In the past, computer vision systems performed well in constrained environments, but not in complex and volatile industrial environments, so many tasks required trained workers to complete. In recent years, computer vision technology based on machine learning has greatly improved the recognition rate of natural image classification in the real world. However, these methods rely on a large amount of training data (millions of training images, which require a long training time (hours to days), so that these machine learning algorithms are not suitable for real-time general-purpose robots. In this paper, we use convolution neural network to recognize objects in a given image. In order to avoid the complicated training for a long time, we input the captured images into the pre-trained convolution neural network and take the output of the convolutional neural network as the feature vector to represent the target in the image. Firstly, the target is tracked by the camera, then the feature vector sequence is stored in memory. Finally, the feature is trained by support vector machine (SVM) to realize the classification of the target. Support vector machine (SVM) can be trained in a short time to meet the real-time requirements. This paper tests the recognition rate of the vision system in the real world indoor environment. With camera and motion detection, the designed system can track objects displayed in the field of vision. If the user starts learning mode, the system will train a new support vector machine for the new object. If the user starts the prediction mode, the system will guess the identity of the object based on the trained model. This paper also studies the possibility of only using visual feedback to classify objects. The system checks the similarity between the Eigenvectors of a given object and any known target Eigenvectors. The similarity is obtained by measuring the Euclidean distance between them. Finally, we use the standard database to test the proposed algorithm and compare it with other databases.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【相似文献】