CASIA OpenIR  > 毕业生  > 硕士学位论文
基于嵌入式视觉系统的小目标实时检测与识别技术研究
谷姣姣
学位类型工学硕士
导师原魁
2017-05-23
学位授予单位中国科学院大学
学位授予地点北京
关键词小目标检测 小目标识别 实时性 Fpga Dsp
中文摘要    目标检测与识别是计算机视觉、人工智能等领域的研究热点,在工业、军事、医疗、安防监控等领域得到广泛应用。然而,随着实际应用的不断增加,人们对目标检测与识别系统的实时性要求也越来越高。虽然近年来目标检测和目标识别算法的研究已经取得了较大进展,但可以同时兼顾实时性和准确性的目标检测和识别应用方面的研究工作仍非常有限,要开发出能够真正满足实时性和鲁棒性要求的目标检测与识别系统还需要解决大量的问题。本文对基于FPGA和DSP的小目标检测和识别方法进行了比较深入的研究,以提高准确率和实时性为目标,解决了多个关键技术问题,为研制开发具有实用价值的嵌入式目标检测和识别系统打下了良好基础。
    本文的主要工作如下:
    1、针对红外图像中车辆小目标的特点,提出了一种基于边缘概率分布与亮度信息相融合的检测方法,并在课题组设计的基于 FPGA 和 DSP 的嵌入式智能图像处理卡上进行了实现。该方法将改进的Sobel算子提取的边缘特征用于朴素贝叶斯分类器进行训练,同时又融合了车辆的亮度信息,用于在上位机上构建训练模型;然后,将构建的模型在嵌入式智能图像卡上进行硬件实现,使其能够满足实时检测系统的需要。
    首先,算法的滤波、边缘检测、二值化、基于朴素贝叶斯模型的车辆目标检测、融合边缘的亮度信息主要在FPGA中进行硬件实现,FPGA同时负责计算目标坐标并将其传递给SRAM。然后,DSP从SRAM中读取目标坐标,根据坐标位置对车辆目标用矩形框进行标注,统计出检测到的目标个数,并将结果送到上位机中进行显示。最后,上位机显示检测结果,并可对FPGA中的参数进行在线调节,参数通过网口传递到DSP,再由DSP传到FPGA中进行实时调整。实验结果表明,上述红外视觉系统不仅可以实时检测出多个车辆小目标,具有较低的时间复杂度和较高的准确率,而且可以检测出被部分遮挡的车辆目标和与相机成一定角度范围内的车辆目标,说明本系统具有较强的鲁棒性和泛化性。
    2、针对战场环境中坦克、飞机、汽车、士兵等小目标的实时识别问题,提出了一种基于小波矩特征与BP神经网络的实时识别方法,并在低功耗的嵌入式智能图像卡上对该方法进行了硬件实现。该方法给出了小波矩特征提取算法的离散化表示,对小波矩特征提取算法和BP神经网络进行了改进,使其能够在DSP中进行定点化实现,降低了时间复杂度。
    为了提高算法的处理速度,本文将图像滤波、自适应阈值计算和图像二值化等图像预处理操作在FPGA中进行硬件实现,而DSP主要完成改进后的小波矩特征提取和BP神经网络分类器的定点化实现。实验结果表明,上述视觉系统不仅能识别单个小目标,也能够同时准确识别出多个小目标;算法具有较低的时间复杂度和平移、旋转、尺度不变性,能够实时识别出发生平移、旋转和尺度变化的小目标。
英文摘要    As a very active research area in the field of computer vision and artificial intelligence, object detection and recognition has been widely used in industrial, military, medical, security monitoring and many other fields. However, with the increasing demands of practical application, the requirement for real-time performance in object detection and recognition system is becoming higher. In recent decades, although researches on object detection and recognition algorithm have made satisfactory progress, researches reconciling the demands for the real-time performance and accuracy at the same time on object detection and recognition are very limited. There are still many problems to be solved in order to build a real-time, robust and practical object detection and recognition system. This dissertation deeply studies the problem on small object detection and recognition using embedded system based on DSP and FPGA, aiming at improving the precision and real-time performance, and solves several key technical problems, which lays a solid foundation for developing real-time object detection and recognition system with practical value.
    The research work of this dissertation can be summarized as the following aspects:
    1. According to the characteristics of vehicles captured by an infrared camera, an infrared small object detection algorithm based on Bayesian edge probability distribution and brightness features is proposed and implemented on an embedded image processing card based on FPGA and DSP, which was designed by our research group. The algorithm puts edge features of the training samples extracted by an improved Sobel operator into the Naive Bayes Model, and then fuses with brightness features to construct the detector template on the host computer. Then, the detector template is hardware implemented on the image acquisition and processing board to meet the requirements of the real-time vehicle detection system.
    Firstly, image filtering, edge extraction, image binaryzation, Naive Bayes based object detection and fusing with brightness features are hardware implemented on FPGA, objects' coordinates are also calculated on FPGA and then passed to the SRAM. Secondly, DSP gets the coordinates from the SRAM, marks vehicle objects with rectangles according to the coordinates, summarizes the total number of detected vehicle objects and then transmits the detection results to PC. Finally, PC displays the detection results and modifies detection parameters of FPGA online, the modified parameters are passed from PC to FPGA through DSP by the 100-Mbps Ethernet interface on the image processing board. Experiment results show that the method can not only detect multiple vehicle objects in real-time with a good computation complexity and accuracy, but also can detect vehicles which are partially occluded and detect vehicles whose left lateral sides have an oblique angle with the camera, which proves that the system has a good robustness and generalization performance.
    2. In order to recognize cars, planes, vehicles, soldiers and other small objects in the battlefield environment, a real-time object recognition method of wavelet moment-based back-propagation(BP) neural network classifier is proposed and hardware implemented on a low-power embedded image acquiring and processing board. The method gives the discrete form of wavelet moment extraction algorithm, improves the wavelet moment extraction algorithm and BP neural network classifier to make a fixed point realization of them on the DSP, and reduces the time complexity.
    To improve the processing speed of the algorithm, image filtering, adaptive threshold calculation and image binaryzation are hardware implemented on FPGA, while the improved wavelet moment extraction algorithm and BP neural network classifier are realized in a fixed point form on the DSP. Experiment results show that the method can recognize single object and multiple objects of similar shape correctly and efficiently. The proposed method is of good computation complexity and has scaling, translation and rotation invariance, and can recognize small objects in real-time correctly even though they have translation, rotation and scaling change.
学科领域控制理论与控制工程
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/14703
专题毕业生_硕士学位论文
作者单位中国科学院自动化研究所
推荐引用方式
GB/T 7714
谷姣姣. 基于嵌入式视觉系统的小目标实时检测与识别技术研究[D]. 北京. 中国科学院大学,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
5_22 原改-谷姣姣毕业论文.pdf(3601KB)学位论文 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[谷姣姣]的文章
百度学术
百度学术中相似的文章
[谷姣姣]的文章
必应学术
必应学术中相似的文章
[谷姣姣]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

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。