基于异构多核处理器的双目视觉系统设计与实现
发布时间：2018-08-26 16:05
【摘要】：双目视觉是计算机视觉领域的重要分支,其通过计算出同一场景下同一空间点在双目图像上的视差来恢复深度信息,这种深度信息恢复方式具有非接触式和被动式两大优点。近年来,随着对双目匹配算法的深入研究,许多算法虽然在准确性上有了很大提高,但它们都以较理想的双目图像为前提。事实上由于现实成像因素的影响,常会导致双目图像在亮度等属性上具有偏差,这就要求算法还应具备较高的鲁棒性,另外,目前准确性较高的算法均具有计算量大、实时性差的特点,使其在要求实时处理的应用中很难运用。SGBM(Semi-global block matching)算法是一种半全局匹配算法,具有较高的准确性,但其鲁棒性与实时性依然有待提高。因此本文对SGBM算法做了改进研究和并行加速方法研究,并将改进后的算法应用到本文双目测距系统中。针对SGBM算法中BT匹配代价对亮度偏差不够鲁棒的问题,提出了一种联合Census匹配代价的BT-Census匹配代价。由于Census匹配代价很好的保留了邻域像素间的结构特性,因此在匹配时提高了对亮度偏差的鲁棒性。为了提高实时性,本文对改进后的算法做了并行加速方法的研究,并基于OpenCL(Open Computing Language)异构并行计算框架做了实现。首先对该算法中匹配代价、代价优化、视差计算及视差精细化模块的主要算法做了并行化分析,然后在OpenCL框架下,对算法的数据存储结构、OpenCL内核做了设计,并结合内核性能评估与算法优化,实现了异构并行加速的算法。实验表明,在同一处理器(AMD APU A8-4555M)上,本文算法在保证准确性相近的前提下,相比于经过SSE2指令集加速的SGBM算法,在实时性能上依然有2.2倍的提升。本文以上述改进的算法为核心,设计与实现了双目测距系统。该双目测距系统,采用了基于异构多核处理器的计算平台,不但减小了多核间的数据传输延迟,而且有效的控制了系统功耗。实验表明,在双目图像尺寸为640×480,最大视差层级为64的条件下,该双目测距系统平均测距周期为110ms,在0.5～5米以内的深度测量误差小于9.6%,初步实现了一套低功耗、近实时的双目测距系统。
[Abstract]:Binocular vision is an important branch in the field of computer vision. The depth information is restored by calculating the parallax of the same space point in the same scene on the binocular image. This depth information restoration method has two advantages: contactless and passive. In recent years, with the in-depth study of binocular matching algorithms, many algorithms have greatly improved in accuracy, but they are based on a more ideal binocular image as the premise. In fact, due to the influence of realistic imaging factors, binocular images often have deviations in luminance and other attributes, which requires that the algorithms should also have higher robustness. In addition, the current algorithms with high accuracy all have a large amount of computation. Because of the poor real-time performance, it is difficult to use SGBM (Semi-global block matching) algorithm, which is a semi-global matching algorithm, in the application of real-time processing, but its robustness and real-time performance still need to be improved. Therefore, the improved SGBM algorithm and the parallel acceleration method are studied in this paper, and the improved algorithm is applied to the binocular ranging system in this paper. Aiming at the problem that the BT matching cost is not robust to the brightness deviation in the SGBM algorithm, a new BT-Census matching cost combining the Census matching cost is proposed. Because the Census match cost is very good to retain the structure characteristics of the neighborhood pixels, so the robustness of the matching algorithm is improved to the luminance deviation. In order to improve real-time performance, the improved algorithm is studied by parallel acceleration method, and implemented based on OpenCL (Open Computing Language) heterogeneous parallel computing framework. Firstly, the main algorithms of matching cost, cost optimization, parallax calculation and parallax refinement module in this algorithm are analyzed in parallel. Then, the data storage structure of the algorithm is designed under OpenCL framework. Combined with kernel performance evaluation and algorithm optimization, heterogeneous parallel acceleration algorithm is realized. Experiments show that on the same processor (AMD APU A8-4555M), the proposed algorithm can improve the real-time performance by 2.2 times compared with the SGBM algorithm accelerated by the SSE2 instruction set, on the premise that the accuracy of the algorithm is similar. In this paper, the binocular ranging system is designed and implemented based on the improved algorithm. The binocular ranging system adopts a computing platform based on heterogeneous multi-core processor, which not only reduces the data transmission delay between multi-cores, but also effectively controls the power consumption of the system. The experimental results show that under the condition that the size of binocular image is 640 脳 480 and the maximum parallax level is 64, the average ranging period of the binocular ranging system is 110 Ms, and the depth measurement error within 0.5 m is less than 9. 6. A set of low power consumption is initially realized. Near-real-time binocular ranging system.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2205431
[Abstract]:Binocular vision is an important branch in the field of computer vision. The depth information is restored by calculating the parallax of the same space point in the same scene on the binocular image. This depth information restoration method has two advantages: contactless and passive. In recent years, with the in-depth study of binocular matching algorithms, many algorithms have greatly improved in accuracy, but they are based on a more ideal binocular image as the premise. In fact, due to the influence of realistic imaging factors, binocular images often have deviations in luminance and other attributes, which requires that the algorithms should also have higher robustness. In addition, the current algorithms with high accuracy all have a large amount of computation. Because of the poor real-time performance, it is difficult to use SGBM (Semi-global block matching) algorithm, which is a semi-global matching algorithm, in the application of real-time processing, but its robustness and real-time performance still need to be improved. Therefore, the improved SGBM algorithm and the parallel acceleration method are studied in this paper, and the improved algorithm is applied to the binocular ranging system in this paper. Aiming at the problem that the BT matching cost is not robust to the brightness deviation in the SGBM algorithm, a new BT-Census matching cost combining the Census matching cost is proposed. Because the Census match cost is very good to retain the structure characteristics of the neighborhood pixels, so the robustness of the matching algorithm is improved to the luminance deviation. In order to improve real-time performance, the improved algorithm is studied by parallel acceleration method, and implemented based on OpenCL (Open Computing Language) heterogeneous parallel computing framework. Firstly, the main algorithms of matching cost, cost optimization, parallax calculation and parallax refinement module in this algorithm are analyzed in parallel. Then, the data storage structure of the algorithm is designed under OpenCL framework. Combined with kernel performance evaluation and algorithm optimization, heterogeneous parallel acceleration algorithm is realized. Experiments show that on the same processor (AMD APU A8-4555M), the proposed algorithm can improve the real-time performance by 2.2 times compared with the SGBM algorithm accelerated by the SSE2 instruction set, on the premise that the accuracy of the algorithm is similar. In this paper, the binocular ranging system is designed and implemented based on the improved algorithm. The binocular ranging system adopts a computing platform based on heterogeneous multi-core processor, which not only reduces the data transmission delay between multi-cores, but also effectively controls the power consumption of the system. The experimental results show that under the condition that the size of binocular image is 640 脳 480 and the maximum parallax level is 64, the average ranging period of the binocular ranging system is 110 Ms, and the depth measurement error within 0.5 m is less than 9. 6. A set of low power consumption is initially realized. Near-real-time binocular ranging system.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前5条
1 邸男;田睿;;空间机器人双目视觉测量系统精度分析[J];载人航天;2017年01期
2 曹之乐;严中红;王洪;;双目立体视觉匹配技术综述[J];重庆理工大学学报(自然科学);2015年02期
3 董峰;孙立宁;汝长海;;基于双目视觉的医疗机器人摆位系统测量方法[J];光电子.激光;2014年05期
4 朱遵尚;苑云;李由;尚洋;于起峰;;嫦娥一号月面成像的高精度匹配及月貌三维重建[J];光学学报;2014年02期
5 于乃功;秦永钢;阮晓钢;;立体匹配算法进展[J];计算机测量与控制;2009年05期
，本文编号：2205431
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2205431.html
