【作者】 李航；

【导师】 向健勇；

【作者基本信息】 西安电子科技大学 ， 物理电子学， 2021， 硕士

【摘要】 目标检测与跟踪作为计算机视觉领域的一个重要分支,在智能交通、安防监控等领域具有广泛应用。目前主流的目标检测与目标跟踪算法虽然具有较高的精度,但对于计算能力相对较低的嵌入式平台,其实时性还不够强。本文主要从目标检测跟踪算法的改进和嵌入式平台的算法移植两个方面展开研究,具体内容如下:(1)本文首先对系统的总体方案进行了规划:在硬件平台方面,本文选取ARM+FPGA架构的SOC(System on Chip)作为系统的核心处理器,并介绍了基于该平台的软硬件协同设计方法、ARM与FPGA之间的高速数据传输方式、HLS(High Level Synthesis)算法开发流程等。在算法设计方面,本文综合考虑了各种算法的优缺点,最终选取与硬件平台具有高匹配度的单帧检测+多帧确认的方案来分别实现目标的检测与跟踪,并分别将这两部分算法部署在FPGA端和ARM端。(2)对于目标检测算法,本文研究了基于自底向上的视觉注意机制原理和基于局部对比度的显著性检测算法。由于常见的基于局部对比度的检测算法只适用于红外目标检测,且在复杂场景下具有误检率高、精度低等缺点,本文提出了一种改进的对比度显著性检测算法。该算法首先通过构建图像金字塔以及扩展检测窗口结构,可有效扩大目标的检测范围;其次,在计算区域显著度时,通过度量多方向局部对比度,可显著提高检测精度,降低误检率,并且使该算法适用于灰度图像的目标检测;然后,该算法根据区域显著度进行非极大值抑制,利用区域生长法进行目标分割,最终实现对候选目标区域的准确定位。仿真结果表明,在同等条件下,本文算法的准确率相较于MPCM(Multiscale Patch based Contrast Measure)算法提高了约8%。(3)针对单帧检测算法提取出的候选目标区域,本文采用帧间匹配的方式完成真伪目标筛选,实现对目标的跟踪。为了增强对目标的特征表示,本文提取了候选目标的大小、位置、灰度均值、灰度值方差、BRISK(Binary Robust Invariant Scalable Keypoints)等多个特征;考虑到多特征匹配时的量纲和权重差异,本算法采用标准化欧式距离进行多特征融合,实现对目标特性的相似性度量;本文还联合时域信息,设计了一套目标置信度评价函数,并对实际场景中可能出现的不同情形进行了模拟仿真,结果表明该评价函数可有效降低误检率。(4)最后,本文利用FPGA并行计算和ARM编程灵活的优势部署检测跟踪算法。本系统采用VDMA(Video Direct Memory Access)实现了ARM与FPGA之间的高速图像数据传输,通过BRAM(Block Random Access Memory)实现算法中间计算结果的传输。利用HLS工具,本文实现了单帧检测算法的流水线设计和并行优化,可以快速筛选出候选目标区域。最后在部署ARM端的算法时,本文对多帧确认算法进行优化和改进,来进一步提高算法的实时性。经过实际测试,本系统可实现多目标的实时检测跟踪,对分辨率为1920×1080的视频图像的处理速度大于30 frame/s,并且当目标短暂消失后,仍然可以重新定位目标,从而实现对目标的准确检测和鲁棒性跟踪。更多还原

【Abstract】 As an important branch in the field of computer vision,target detection and tracking have a wide range of applications in intelligent transportation,security monitoring and other fields.Although the current mainstream target detection and target tracking algorithms have high accuracy,their real-time performance is not strong enough for embedded platforms with relatively low computing power.This article focuses on the improvement of the target detection and tracking algorithm and the transplantation of the embedded platform algorithm.The specific contents are as follows:(1)This thesis first plans the overall scheme of the system: In terms of hardware platform,this thesis selects the ARM+FPGA architecture SOC(System on Chip)as the core processor of the system,and introduces the software and hardware co-design,high-speed data transmission between ARM and FPGA,HLS(High Level Synthesis)algorithm development process,etc.In terms of algorithm design,considering the advantages and disadvantages of various algorithms,this thesis selects a single-frame detection + multi-frame confirmation scheme with a high degree of matching with the hardware platform to achieve target detection and tracking,and deploy the two parts of the algorithm on the FPGA side and the ARM side respectively.(2)For the target detection algorithm,this paper studies the principle of bottom-up visual attention mechanism and the saliency detection algorithm based on local contrast.Since the common detection algorithm based on local contrast is only suitable for infrared target detection,and has the shortcomings of high false detection rate and low accuracy in complex scenes,this thesis proposes an improved contrast saliency detection algorithm.The algorithm firstly builds an image pyramid and expands the detection window structure to expand the detection range of the target effectively.Secondly,when calculating the saliency of the area,by measuring the local contrast in multiple directions,the algorithm can significantly improve the detection accuracy,reduce the false detection rate,and is suitable for target detection in grayscale images.Then,the algorithm performs non-maximum suppression according to the saliency of the region,and uses the region growing method to segment the target,and finally realizes the accurate positioning of the candidate target region.Simulation results show that the performance of this algorithm is better than other detection algorithms based on local contrast.(3)Aiming at the candidate target area extracted by the single-frame detection algorithm,this thesis uses the method of inter-frame matching to complete the screening of true and false targets,and realize the tracking of the target.In order to enhance the feature representation of the target,this thesis extracts multiple features such as the size,position,gray-scale mean,gray-value variance,and BRISK(Binary Robust Invariant Scalable Keypoints)of the candidate target.Taking into account the difference of dimension and weight in multi-feature matching,this algorithm uses standardized Euclidean distance for multi-feature fusion to achieve similarity measurement of target features.This thesis also combines time-domain information to design a set of target confidence evaluation function,and simulates the different situations that may appear in the actual scene.The results show that the evaluation function can further reduce the false detection rate.(4)Finally,this thesis uses the advantages of FPGA parallel computing and ARM programming flexibility to deploy detection and tracking algorithms.This system uses VDMA(Video Direct Memory Access)to achieve high-speed image data transmission between ARM and FPGA,and uses BRAM(Block Random Access Memory)to achieve the transmission of intermediate calculation results of the algorithm.Using HLS tools,this thesis realizes the pipeline design and parallel optimization of the single-frame detection algorithm,which can quickly screen out candidate target regions.Finally,when deploying the ARMside algorithm,this thesis optimizes and improves the multi-frame confirmation algorithm to further improve the real-time performance of the algorithm.After the actual test,the system can track the real-time detection of multiple targets,the processing speed of this system for video images with a resolution of 1920×1080 is greater than 30 frame/s,and when the target disappears briefly,the system can still relocate the target,so as to achieve accurate detection and robust tracking of the target.更多还原