【作者】 孟祥杰；

【导师】 朱齐丹；

【作者基本信息】 哈尔滨工程大学 ， 控制理论与控制工程， 2011， 硕士

【摘要】 行人检测是当前计算机视觉和人工智能领域研究的重点和热点,也是目标检测的重要分支,其在军事领域、智能交通、机器人导航、智能监控、人体运动分析等领域都有广泛的应用前景。由于人体的衣着、姿势、光照各异,行人检测是非常具有挑战性的课题。目前,计算机视觉领域的主流思想将行人检测问题认为是一个模式分类问题,其方法是从大量的行人训练样本中提取特征,利用机器学习方法将行人与其他运动目标以及干扰背景区分开来,并准确定位。全景视觉系统具有视场范围大的特点,当前已经广泛地应用在机器人导航、空间探测、视频监控、虚拟现实、环境感知技术等领域。行人检测技术和全景视觉系统的结合更能充分发挥各自的优势,提供更加广泛的应用前景。在全景视觉中研究行人检测,可以充分利用全景视觉视野范围大的特点,检测到360°视野范围内的行人,更好的满足机器人导航和智能监控等应用场合的需求。近几年来,GPU和基于GPU的通用计算得到了迅速的发展,基于GPU的通用计算在计算机视觉领域得到广泛的应用。本文主要研究基于双曲面折反射全景视觉系统的行人检测方法,采用积分梯度方向直方图特征和线性支持向量机设计了基于全景视觉的行人检测器,并在GPU平台上利用并行计算技术改进了行人检测算法的关键步骤,在保持检测准确率的前提下,获得明显的检测速度上的提升。首先研究了双曲面折反射全景视觉的成像机理,推导了双曲面折反射成像系统的数学模型。为了克服全景图像畸变对行人检测的影响,对原始全景图像进行还原解算,从而获得符合正常人的视觉感受的全景图像,保证全景视觉中行人检测的准确性。然后本文设计了一个基于全景视觉的行人检测器,采用Dalal提出的梯度方向直方图特征为基础,应用了用积分图像方法来计算HOG特征集的方法,大大提高了HOG特征的计算速度,同时采用线性支持向量机作为行人分类器,对在行人检测中存在大量的重叠结果窗口,采用非极大值抑制算法进行多结果融合,获得较好的检测效果。最后本文在GPU平台上对基于全景视觉的行人检测器进行了改进,采用CUDA并行计算技术加速全景图像还原解算,积分梯度方向直方图计算和线性支持向量机检测。本文对算法的关键步骤所需时间以及资源占用情况进行了详细的分析。实验结果表明,该方法在保持行人检测准确性的同时,极大地提升了在全景图像中行人检测的速度,我们的算法所需时间只是在CPU上实现的算法的1/8左右。更多还原

【Abstract】 Nowadays, pedestrian detection is intensively investigated and becoming a hot topic in the fields of computer vision and artificial intelligence. It’s also an important branch of object detection. It could be widely used in the military field,intelligent transportation systems,robot navigation, intelligent surveillance, human motion analysis and so on. Detecting human in images is a challenging task because of the variability in clothing and illumination conditions, and the wide range of poses that people can adopt。Nowadays, pedestrian detection is considered as a pattern classification problem in the filed of computer vision. The Algorithm mainly based on machine learning, which extracts features from pedestrian samples and distinguish pedestrian from other targets and background area, to find the accurate location of pedestrian.Panoramic vision system has a large field of view.It’s widely used in robot navigation, space probe,video surveillance, virtual reality, environment sensing technology and so on. The combination of pedestrian detection technology and panoramic vision system could play their respective advantages.It’s potential application is very promising.Pedestrian detection based on panoramic vision system could take advantage of a large panoramic field of view and detect pedestrian in the 360-degree field of view, to better meet needs of robot navigation intelligent surveillance and other applications. Last few years, Graphics processing unit and general-purpose computing on GPU has been in rapid development,general-purpose computing on graphics processing units has been widely used in computer vision.This paper mainly studies the pedestrian detection algorithm based on panoramic vision system, designs a pedestrian detector based on panoramic vision with integral histograms of oriented gradient features and linear support vector machine, and improves the key steps of the pedestrian detection algorithm with parallel computing technology on GPU platform, achieves significant improvement on the detection speed, while maintaining the premise of detection accuracy。Firstly, by researching the imaging principle of panoramic vision, derive mathematical model of the hyperboloid reflection imaging system. In order to overcome the impact of panoramic image distortion on pedestrian detection,a cylinder unwarping algorithm based on geometry principle of panoramic image is proposed can obtain 360-dcgrce panoramic image. Ensure the detection accuracy of pedestrian detection in panoramic image.Secondly, this paper designs a pedestrian detector based on panoramic vision, based on histogram of oriented gradient features proposed by Dalal, Introduced with the integral image method to calculate the HOG feature set, greatly improved the calculation speed of HOG features. At the same time using a linear support vector machine as a pedestrian classifier. In the pedestrian detection, there are a large number of overlapping windows, we use non-maxima suppression algorithm for integration of multiple results,to obtain better detection results.Finally, this paper improved the pedestrian detector based on panoramic vision on GPU platform,using CUDA parallel computing technology to accelerate to obtain panoramic image, calculate the integral histogram of oriented gradient and the linear support vector machine detection. we detailed analyzes the processing times and the occupancy of each kernel used by our algorithm. Experimental results show that our method achieves significant improvement on the detection speed in panoramic image, and we report a speed up by a factor of 8 over our CPU implementation, while maintaining the premise of detection accuracy.更多还原