【作者】 王萌；

【导师】 王酉； 郝杰；

【作者基本信息】 浙江大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 随着时代发展,移动机器人的相关技术发展迅速。球形机器人由于其较好的密封性、运动中的抗倾倒性在危险环境探测、安防巡检等方面的工作上拥有着得天独厚的优势。执行巡检工作时,需要能够有效的检测移动目标,进而执行下一步的避障指令,因此需要一套针对球形机器人的感知方案。由于球形机器人特殊的结构特征与运动特点,研究的主要挑战在于由于滚动导致的图像质量不佳影响检测效果问题以及由于计算资源有限,设计算法需要平衡算法的计算量与检测效果的问题,本文的主要工作如下:1.由于实际使用传感器数据与公开数据集数据存在差异,为了验证算法效果,对目标检测系统的传感器进行标定后,在不同时间与场景采集处理了机器人专用的目标检测数据集,一共7122帧。2.设计并且搭建了球形机器人的目标检测系统。选用毫米波雷达与工业相机作为检测目标的传感器,惯性测量单元（Inertial Measurement Unit,IMU）作为数据校正的辅助传感器,分析选择了目标检测系统的安装方案。3.针对球形机器人的目标检测系统提出了配套的目标检测算法。在特征提取部分基于Mobilenet V3融合毫米波雷达数据,提升了目标检测的准确度。在nu Scenes数据集上,算法m AP₅₀达到了74.33%,检测速率约为7.66帧/秒,在速率与精确率上达到了一个较好的平衡,最终输出结构化的检测结果数据。4.针对单帧图像检测结果不稳定现象提出了基于连续帧的检测优化模块。使用自适应的交互式多模型算法进行运动建模,并且通过扩展卡尔曼滤波器实现多帧数据的连接,增加信息量。通过级联匹配关联图像与毫米波雷达数据并融合,改善了检测效果。更多还原

【Abstract】 With the development of the times,the related technologies of mobile robots are developing rapidly.Spherical robots have unique advantages in dangerous environment detection and security inspection due to their good sealing performance,anti-interference performance and anti-dumping performance during motion.When it is working,how to detect moving targets effectively,and then avoid obstacles requires us to design appropriate perception algorithms.Due to the special structural features and motion characteristics of the spherical robot,the main challenge of the research is that due to the rolling motion,the image quality is poor and due to the shortage of computing resources,the performance of the perception algorithm needs to be better on a lightweight basis.The main work is as follows:1.Due to the difference between the actual sensor data and the public data set, in order to better verify the effectiveness of the algorithm,after independent calibration and joint calibration of the sensor,the spherical robot target detection data was collected at different times and scenes.We collected a total of 7122 images and their corresponding other sensor data,labeled the images,to verify the experimental results.2.A target detection system for a spherical robot is designed and built.We choose millimeter-wave radar and camera as the sensor to detect the target,and inertial measurement unit as the auxiliary sensor for data correction.According to the structure of the spherical robot,the installation scheme of the target detection system is selected.3.A set of lightweight object detection algorithms suitable for spherical robots is designed.This algorithm is based on the design of the lightweight feature extraction network Mobilenet V3,and integrates the data of the millimeter-wave radar to improve its accuracy.In the nu Scenes dataset,the algorithm m AP₅₀ reaches 74.33%,and the detection rate is about 7.66fps,achieving a good balance.The algorithm finally outputs structured detection result data.4.Due to the limited single-frame data,the detection results are unstable,and missed detections may occur.We propose an object tracking based detection optimization module.We predict the target state by using the extended Kalman filter,and adaptively switch the target motion model by using interactive multi-model.We associate image and millimeter-wave radar data through cascade matching and design a multi-sensor decision-level fusion scheme to effectively improve the detection effect.更多还原