基于雷达和视觉融合的行人检测方法研究
文献类型：学位论文
| 作者 | 郭晓鹏; 马亚雄
|
| 学位类别 | 硕士 |
| 答辩日期 | 2018-05-17 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 授予地点 | 沈阳 |
| 导师 | 杜劲松 |
| 关键词 | 毫米波雷达 视觉 数据融合 行人检测 似物性 |
| 其他题名 | Research of pedestrian detection based on fusion of radar and vision |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 本文提出将毫米波雷达和视觉进行融合来实现行人检测，以达到同时获取行人身份信息和方位信息的目的。论文主要研究工作如下：研究并分析毫米波雷达下行人信号特性，针对低检测阈值状态下行人信号断续和环境噪声问题，设计了基于帧内聚类和帧间跟踪的行人信号检测算法，以提取有效行人信号。学习图像识别机制和原理，分析了各种行人识别特征的特点。针对传统滑动窗口检测算法效率低下的问题，将似物性检测引入图像识别过程，并加以改进，实现先提取图像中潜在目标区域，然后提取其梯度方向直方图特征，最后用支持向量机分类器对其分类以实现视觉行人快速检测。为实现雷达和视觉数据的融合，研究了相机成像模型和标定技术，建立雷达-相机坐标系，并利用最小二乘拟合法求取空间转换矩阵，实现了数据空间对准。然后根据雷达数据的图像投影点和目标距离信息，自动确定潜在目标区域位置及其大小。针对拟合偏差所导致的漏检现象，提出一种渐进式动态搜索潜在目标区域的策略。实验表明，该融合机制可以有效地确定视觉候选目标区域。最后，对上述融合算法所确定的视觉潜在目标区域，应用所提出的改进后的视觉检测算法，实现了行人检测并获取其方位信息。 |
| 英文摘要 | This paper proposes a method for pedestrian detection based on fusion of millimeter wave radar and vision to realize pedestrian detection and obtain its relative dynamic information. The main research work is as following: The characters of pedestrian signal detected by millimeter wave radar is studied and analysed.To deal with the discontinuous of pedestrian radar signal and environment noise problem caused by low detection threshold, a algorithm based on near neighbor intra-frame cluster and inter-frames track is designed to extract valid pedestrian signal. Learn the mechanism and principle of image recognition, several kinds pedestrian recognition features are introduced and analysed. To deal with the low detection efficiency caused by traditional sliding window detection algorithm, Objectness estimation is applied to the visual detection process and some improvement is made. This algorithm is able to produce some candidate boxes which maybe contain a potential target. At last, HOG features are extracted from these candidate boxes and support vector machine is employed to realize the final pedestrian recognition quickly. For space-time fusion of radar and vision data, the camera image model and calibration technology are studied. The radar-camera coordinate system is established. Least-squares fitting method is adopted to obtain the space transformation matrix.Then space alignment target is achieved through transformation matrix. Based on the projection of radar points on image and target’s dynamic infomation, potential target area of image and its size are determined automatically. In addition, to cope with phenomenon of miss detection caused by the fitting deviation, a progressive dynamic search strategy for the region of interest is proposed. Experiment results show that this fusion mechanism can effectively capture the visual candidate target area. Finally, the proposed visual recognition algorithm is applied to the potential target area produced by above space matching algorithm to realize the final task, experimental results show that the proposed fusion detection algorithm has the ability to detect human and acquire its position information. |
| 语种 | 中文 |
| 产权排序 | 1 |
| 页码 | 68页 |
| 源URL | [http://ir.sia.cn/handle/173321/21754]  |
| 专题 | 沈阳自动化研究所_装备制造技术研究室 |
| 推荐引用方式 GB/T 7714 | 郭晓鹏,马亚雄. 基于雷达和视觉融合的行人检测方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所. 2018. |
入库方式： OAI收割
来源：沈阳自动化研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。