基于多传感器融合的机器人目标物标定

发布时间：2018-12-17 02:09

【摘要】：随着传感器技术和计算机科学的迅速发展,人们广泛利用机器人在危险的环境或人类难以到达的地方执行任务。移动机器人本身具有传感器和处理器,可以自主进行探测、判断、决策并在到达目标地点后执行预先设定的任务。但由于未知环境的不确定性,单一的传感器难以满足复杂任务的要求,多传感器数据融合技术能够综合各个传感器的信息,进而获得更全面、更准确的决策信息。本文利用多传感器数据融合技术,设计并开发了一套基于Khepera IV嵌入式机器人在未知环境下的执行目标标定任务系统。机器人在未知环境中移动,避开障碍物,并运用自带的摄像头和超声波传感器寻找目标物。首先对不同角度拍摄的目标物的图片进行预处理,提取出目标物的特征,存入上位机。机器人在运动的过程中,运用超声波传感器检测周围的物体,并拍照将照片上传到上位机与目标物的特征比对,判断是否是目标物。如果找到目标物,则标定出目标物的位置。研究内容包括以下几个方面:1.利用zabbix开源软件,在windows上位机的虚拟机中搭建机器人监控系统。对机器的实时状态进行监控,包括机器人电池、CPU占用率和机器人的运动轨迹等,存放在MySQL数据库中,便于在机器人的运动控制和数据融合的过程中的使用。同时,zabbix软件中基于web的数据交互页面可以直观的显示被监控数据的变化情况。2.建立机器人的运动模型,并结合Khepera IV机器人的硬件配置,分析速度控制、位置控制、方向控制的方法,提出了在机器人在运动过程中的模糊避障策略,并对避障策略进行了仿真,仿真结果表明所提出的模糊避障策略是有效的。3.提出了基于机器学习和图像处理的目标物识别方法。首先对目标物进行预处理,提取目标物的特征。在识别之前,运用机器人的摄像头在不同距离、不同视角拍摄目标物的图片,以200张含有目标物的图像作为正样本,300张不含有目标物的图像作为负样本,使用HOG特征提取算法提取特征,对正负样本进行标记,利用线性SVM分类器进行训练,得到一个二值分类器。然后在机器人搜索目标物的过程中,拍摄周围物体,传回上位机进行处理,通过图像增强、二值化、边缘检测、霍夫变换等预处理,在图片中划分出可能存在目标物的区域,将此区域用SVM分类器进行识别,判断是否为目标物。并估算目标物的实际大小,估计目标物的位置。4.进行机器人目标标定的实验,进行机器人监控系统的搭建、实现数据传输和实时图像处理。实验表明,机器人可以实现在未知环境中(障碍物比较规则)的目标识别任务,说明了本文所设计系统的有效性。
[Abstract]:With the rapid development of sensor technology and computer science, robots are widely used to perform tasks in dangerous environments or places that are difficult to reach by human beings. The mobile robot itself has sensors and processors, which can detect, judge, make decisions and perform predefined tasks after reaching the target location. However, because of the uncertainty of unknown environment, it is difficult for a single sensor to meet the requirements of complex tasks. The multi-sensor data fusion technology can synthesize the information of each sensor and obtain more comprehensive and accurate decision information. In this paper, we design and develop a target calibration task system based on Khepera IV embedded robot in unknown environment by using multi-sensor data fusion technology. The robot moves in unknown environment, avoids obstacles, and uses its own camera and ultrasonic sensor to find objects. Firstly, the images of the objects taken from different angles are preprocessed to extract the features of the objects and stored in the upper computer. In the process of motion, the robot uses ultrasonic sensor to detect the objects around it, and takes pictures and compares the features of the object with the host computer to judge whether the object is the object or not. If the object is found, the location of the object is calibrated. The research includes the following aspects: 1. Using zabbix open source software, the robot monitoring system is built in the virtual machine of windows host computer. The real-time state of the machine is monitored, including the battery of the robot, the CPU occupancy rate and the trajectory of the robot. It is stored in the MySQL database, which is convenient for use in the process of robot motion control and data fusion. At the same time, the zabbix software based on web data exchange page can visually display the monitored data changes. 2. The motion model of the robot is established, and the method of speed control, position control and direction control is analyzed based on the hardware configuration of the Khepera IV robot, and the fuzzy obstacle avoidance strategy is put forward in the course of the robot movement. The simulation results show that the proposed fuzzy obstacle avoidance strategy is effective. A method of object recognition based on machine learning and image processing is proposed. First, the target is pretreated to extract the characteristics of the target. Prior to recognition, the robot's camera was used to take pictures of the object from different angles of view at different distances. 200 images containing the object were used as positive samples, and 300 images without the object were taken as negative samples. HOG feature extraction algorithm is used to extract features, positive and negative samples are marked, and linear SVM classifiers are used to train them to obtain a binary classifier. Then, in the process of robot searching for objects, the objects around them are photographed and sent back to the upper computer for processing. Through the preprocessing of image enhancement, binarization, edge detection and Hoff transform, the region where the object may exist is divided in the picture. This area is identified by SVM classifier to determine whether it is the target. And estimate the actual size of the object, estimate the location of the target. 4. The experiment of robot target calibration is carried out, and the robot monitoring system is built to realize data transmission and real time image processing. Experiments show that the robot can realize the target recognition task in unknown environment (obstacle comparison rules), and the effectiveness of the system designed in this paper is illustrated.
【学位授予单位】：东华大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP242

【参考文献】