模糊视频中车辆多目标跟踪及CUDA加速

发布时间：2018-03-14 19:52

本文选题：图像增强　切入点：自适应　出处：《大连海事大学》2017年硕士论文　论文类型：学位论文

【摘要】：在计算机视觉领域,目标跟踪扮演者很重要的角色。优秀的跟踪算法可用于在监控视频中对人流车辆的数量统计,甚至可用于导弹跟踪,拦截不法入侵。为了提升在恶劣环境中对车辆目标进行跟踪的性能,本文通过对雨雪、雾霾等物理模型以及在这些场景下所拍摄图像的特点深入研究,在Stuck跟踪算法的基础之上,提出了一种恶劣场景下的快速自适应跟踪算法,并将此算法应用到多目标的跟踪过程。本文的主要工作如下:在雨雪、雾霾天气下所拍摄的图像会出现降质,对比度低,模糊等现象,本文通过实验分别对比使用多尺度Retinex、暗通道先验去雾算法以及优化的对比度增强算法进行图像增强后的效果,并分别统计不同场景及分辨率下的算法耗时。为了解决跟踪过程中车辆目标的尺度变化并结合原始Struck算法中生成候选样本的方式,本文提出了基于多尺度因子数组的自适应跟踪算法。通过调节数组中尺度因子的数量和变化的范围,可以生成不同数量和尺度的候选样本,使跟踪窗能够自适应车辆大小,提高跟踪的精度。虽然改进后的自适应跟踪算法能够提高跟踪的精度,但是对数量庞大的候选样本进行特征计算、分类器评估与更新等操作又降低了算法的实时性,为了将此算法应用到多目标跟踪的场景中,因此本文基于CUDA架构对算法进行并行优化,将耗时的过程放置在GPU上运行。通过对原始算法的各个模块的并行优化与改进,算法的运行速度较之前提升了三倍。针对原始算法仅适用于单目标以及对遮挡处理的鲁棒性较低的问题,本文通过初始化多个跟踪器并加入多线程实现了多个目标的跟踪,使得Struck算法能够得以更广泛的应用。使用改进后的算法在公共车辆图像数据集上进行测试,跟踪的结果表明,改进后的算法能够解决跟踪过程中车辆目标的尺度变化以及短暂的目标遮挡问题。相对于原始目标跟踪框架,改进后的Struck算法在跟踪精度和速度上都有了很大提高。
[Abstract]:In the field of computer vision, target tracking plays an important role. Excellent tracking algorithms can be used to measure the number of people and vehicles in surveillance videos, and even to track missiles. Intercepting illegal intrusion. In order to improve the performance of tracking vehicle targets in harsh environments, this paper studies the physical models such as rain and snow, haze and the characteristics of images taken in these scenes. Based on the Stuck tracking algorithm, this paper proposes a fast adaptive tracking algorithm for bad scenes, and applies the algorithm to the tracking process of multiple targets. The main work of this paper is as follows: in the rain and snow, The images taken in haze weather will appear the phenomena of degradation, low contrast, blur and so on. In this paper, we compare the effect of image enhancement by using multi-scale Retinex, dark channel prior de-fogging algorithm and optimized contrast enhancement algorithm, respectively. In order to solve the scale change of vehicle target in tracking process and combine with the method of generating candidate samples in original Struck algorithm, the algorithm is used to calculate the time consuming of different scene and resolution. In this paper, an adaptive tracking algorithm based on multi-scale factor array is proposed. By adjusting the number and range of scale factors in the array, candidate samples with different numbers and scales can be generated, so that the tracking window can adapt to the vehicle size. Although the improved adaptive tracking algorithm can improve the tracking accuracy, but the feature calculation of a large number of candidate samples, classifier evaluation and update operations also reduce the real-time performance of the algorithm. In order to apply this algorithm to the scene of multi-target tracking, this paper optimizes the algorithm in parallel based on CUDA architecture and runs the time-consuming process on GPU. The running speed of the algorithm is three times faster than before. Aiming at the problem that the original algorithm is only suitable for single object and low robustness to occlusion processing, this paper initializes multiple trackers and adds multiple threads to realize the tracking of multiple targets. Struck algorithm can be more widely used. The improved algorithm is used to test the image data set of public vehicles. The tracking results show that, Compared with the original target tracking framework, the improved Struck algorithm can solve the problem of vehicle target scale change and temporary object occlusion. Compared with the original target tracking framework, the improved Struck algorithm can greatly improve the tracking accuracy and speed.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】