基于多特征融合与分层数据关联的空中红外多目标跟踪方法

针对星空背景下目标相似度高、数量大和误检数目较多所导致的空中红外多目标跟踪困难问题,提出基于分层数据关联的空中红外多目标在线跟踪方法。首先,根据红外场景特性来提取目标的位置特征、灰度特征和尺度特征;其次,综合这三个特征来计算目标与轨迹之间的初步关联关系以获得真实目标;再次,将所获得的真实目标按照尺度大小分类,大尺度类目标数据关联采用表观特征、运动特征、尺度特征三种特征相加的方法来计算,小尺度类目标数据关联采用表观特征与运动特征两种特征相乘的方法来计算;最后,根据匈牙利算法对两类目标分别进行目标分配、完成轨迹更新。多种复杂情况下的实验结果表明：与仅采用运动特征的在线跟踪方法相比,所提方法的跟踪准确率提升了12.6%;与采用多特征融合的方法相比,所提方法的分层数据关联不仅提高了跟踪速度,也使跟踪准确率提升了19.6%。综上,该方法不仅跟踪精度高,而且具有较好的实时性和抗干扰能力。     

基于多特征融合与分层数据关联的空中红外多目标跟踪方法

针对星空背景下目标相似度高、数量大和误检数目较多所导致的空中红外多目标跟踪困难问题，提出基于分层数据关联的空中红外多目标在线跟踪方法。首先，根据红外场景特性来提取目标的位置特征、灰度特征和尺度特征；其次，综合这三个特征来计算目标与轨迹之间的初步关联关系以获得真实目标；再次，将所获得的真实目标按照尺度大小分类，大尺度类目标数据关联采用表观特征、运动特征、尺度特征三种特征相加的方法来计算，小尺度类目标数据关联采用表观特征与运动特征两种特征相乘的方法来计算；最后，根据匈牙利算法对两类目标分别进行目标分配、完成轨迹更新。多种复杂情况下的实验结果表明：与仅采用运动特征的在线跟踪方法相比，所提方法的跟踪准确率提升了12.6%；与采用多特征融合的方法相比，所提方法的分层数据关联不仅提高了跟踪速度，也使跟踪准确率提升了19.6%。综上，该方法不仅跟踪精度高，而且具有较好的实时性和抗干扰能力。     

基于改进霍夫森林框架的多目标跟踪算法

针对单目视觉对多个相似的目标跟踪因遮挡等因素影响而失效的问题,提出一种基于改进霍夫森林框架的多目标跟踪算法。在将多目标跟踪问题归结为基于目标检测的轨迹关联过程基础上,通过引入在线学习霍夫森林框架将轨迹关联计算转化为最大后验概率（MAP）问题。通过在线采集多目标样本、提取目标外观和运动特征构建霍夫森林,进行森林训练得到轨迹关联概率,从而关联多目标轨迹;而引入低秩逼近Hankel矩阵进行轨迹校验,修复了误匹配的轨迹,改进了在线更新训练样本算法的效能。实验表明,轨迹误匹配率显著改善,能有效提高单目摄像机对多个相似目标有遮挡情况下跟踪的准确性和鲁棒性。     

自适应在线判别外观学习的分层关联多目标跟踪

目的 复杂场景下目标频繁且长时间的遮挡、跟踪目标外观相似引起身份转换等问题给多目标跟踪带来许多挑战。针对多目标跟踪在复杂场景中因长时间遮挡引起身份转换和轨迹分段的问题,提出一种基于自适应在线判别外观学习的分层关联多目标跟踪算法。方法 利用轨迹置信度将多目标跟踪分为局部关联和全局关联两个层次。在局部关联中,置信度高的可靠轨迹利用外观、位置-大小相似度与当前帧检测点进行关联;在全局关联中,置信度低的不可靠轨迹引入运动模型和有效关联范围进一步关联分段的轨迹。在提取目标外观特征时引入增量线性可判别分析方法以解决身份转换问题,依据新增样本与目标样本均值的外观特征差异自适应地更新目标外观模型。结果 在公开数据集2D MOT2015中的PETS09-S2L1、TUD-Stadmitte、Town-Center 3个数据集中与当前10种多目标跟踪算法进行比较,该方法对各个数据集身份转换和轨迹分段都有减少,其中在Town-Center数据集中,身份转换减少了60个,轨迹分段减少了84个,跟踪准确度提高了5.2%以上。结论 本文多目标跟踪方法,能够在复杂场景中稳定有效地实现多目标跟踪,减少轨迹分段现象,其中引入的在线线性可判别外观学习对遮挡产生的身份转换具有良好的解决效果。     

Efficient detection and tracking of moving objects in geo-coordinates

We present a system to detect and track moving objects from an airborne platform. Given a global map, such as a satellite image, our approach can locate and track the targets in geo-coordinates, namely longitude and latitude obtained from geo-registration. A motion model in geo-coordinates is more physically meaningful than the one in image coordinates. We propose to use a two-step geo-registration approach to stitch images acquired by satellite and UAV cameras. Mutual information is used to find correspondences between these two very different modalities. After motion segmentation and geo-registration, tracking is performed in a hierarchical manner: at the temporally local level, moving image blobs extracted by motion segmentation are associated into tracklets; at the global level, tracklets are linked by their appearance and spatio-temporal consistency on the global map. To achieve efficient time performance, graphics processing unit techniques are applied in the geo-registration and motion detection modules, which are the bottleneck of the whole system. Experiments show that our method can efficiently deal with long term occlusion and segmented tracks even when targets fall out the field of view.     

融合SPA遮挡分割的多目标跟踪方法

复杂环境下的多目标视频跟踪是计算机视觉领域的一个难点,有效处理目标间遮挡是解决多目标跟踪问题的关键。将运动分割方法引入目标跟踪领域,提出一种融合骨架点指派(SPA)遮挡分割的多目标跟踪方法。由底层光流信息得到骨架点,并估计骨架点遮挡状态;综合使用目标外观、运动、颜色信息等高级语义信息,将骨架点指派给各个目标;最后以骨架点为核,对运动前景密集分类,得到准确的目标前景像素;在粒子滤波器跟踪框架下,使用概率外观模型进行多目标跟踪。在PETS2009数据集上的实验结果表明,文中方法能够改进现有多目标跟踪方法对目标间交互适应性较差的缺点,更好地处理动态遮挡问题。     

基于DenseNet的无人机在线多目标跟踪算法

多目标跟踪任务的目的,是对图像序列中不同的目标设置不同的编号(ID),最终得到不同目标的运动轨迹。本文针对跟踪过程中目标ID极易变化的现象,提出了一种新的在线多目标跟踪算法。算法主要包含三个步骤:输入预处理、特征提取和数据关联。其中预处理步骤使用NMS算法对输入的检测结果进行筛选;特征提取步骤使用密集连接的特征提取网络对目标进行外观特征的提取,输出特征向量矩阵;数据关联步骤则使用级联匹配的方式,依据目标的位置信息和外观特征信息为其分配各自的ID。此外,该文还整理了一个具有挑战性的无人机场景下的多目标跟踪测试集。实验结果表明,该方法有效地减少了错误的目标ID变化,提高了多目标跟踪算法面对复杂场景时的精度,并保持较快的运行速度。     

Multi-Target Tracking by Online Learning a CRF Model of Appearance and Motion Patterns

We introduce an online learning approach for multi-target tracking. Detection responses are gradually associated into tracklets in multiple levels to produce final tracks. Unlike most previous approaches which only focus on producing discriminative motion and appearance models for all targets, we further consider discriminative features for distinguishing difficult pairs of targets. The tracking problem is formulated using an online learned CRF model, and is transformed into an energy minimization problem. The energy functions include a set of unary functions that are based on motion and appearance models for discriminating all targets, as well as a set of pairwise functions that are based on models for differentiating corresponding pairs of tracklets. The online CRF approach is more powerful at distinguishing spatially close targets with similar appearances, as well as in tracking targets in presence of camera motions. An efficient algorithm is introduced for finding an association with low energy cost. We present results on four public data sets, and show significant improvements compared with several state-of-art methods.     

基于深度学习的分层关联多行人跟踪

由于连续帧之间数据关联的不确定性和所提取外观特征的鉴别力不足,多目标跟踪容易受目标外观变化、运动状态变化、相似目标以及目标消失再出现等干扰因素的影响,出现轨迹ID变换的问题,从而限制基于轨迹分析的行为检测、姿态识别等研究的性能。为了提高数据关联的可靠性从而减少轨迹ID变换,提出了一种基于轨迹置信度的分层数据关联方式。同时,为了提高用于数据关联的外观特征的鉴别力,提出融合了过程层的特征提取网络。在公开的MOT16测试数据集上验证表明,该算法在保证跟踪准确度和精确度的同时,有效降低了轨迹ID变换的次数,提高了跟踪性能。     

Multi-object tracking via MHT with multiple information fusion in surveillance video

Tracking multiple objects is critical to automatic video content analysis and virtual reality. The major problem is how to solve data association problem when ambiguous measurements are caused by objects in close proximity. To tackle this problem, we propose a multiple information fusion-based multiple hypotheses tracking algorithm integrated with appearance feature, local motion pattern feature and repulsion–inertia model for multi-object tracking. Appearance model based on HSV–local binary patterns histogram and local motion pattern based on optical flow are adopted to describe objects. A likelihood calculation framework is proposed to incorporate the similarities of appearance, dynamic process and local motion pattern. To consider the changes in appearance and motion pattern over time, we make use of an effective template updating strategy for each object. In addition, a repulsion–inertia model is adopted to explore more useful information from ambiguous detections. Experimental results show that the proposed approach generates better trajectories with less missing objects and identity switches.     

视频识别中基于簇的在线运动分割算法研究

仿射相机模型下,运动分割问题转化为子空间分离问题,处理这类问题的算法大多是离线算法,当假设不满足时性能很不理想.针对上述问题,提出一种在线运动分割算法,通过动态标签传输和簇分割进行运动分割.首先,根据固定数量的帧进行初始化,接着,通过在线策略更新轨迹相似性,最后,利用动态标签传输技术在帧间传输信息,对簇进行评估和归一化切分成本估计,实现动态的簇分割.基于基准数据集的仿真实验结果表明,算法的运行结果与离线算法相当.     

多传感器广义概率数据关联算法研究

随着跟踪环境、跟踪对象的不断变化与发展,目标与量测已经很难用一一对应的关系去描述;广义数据关联算法（GPDA）提出目标与量测多多对应的可行性规则,其性能在目标与量测无论是否在一一对应的情况下,均优于传统的JPDA算法,并且计算量、存储量均远小于JPDA;考虑到GPDA的上述优点,将GPDA算法拓展到多传感器数据关联,提出了多传感器广义概率数据关联算法（MSGPDA）来处理多传感器数据关联问题;仿真表明,MSGPDA算法由于利用多传感器信息,性能得到明显提高。     

基于运动检测的多车辆跟踪方法研究

针对交通监控场景中多目标粘连造成跟踪上的困难和前后两帧车辆关联困难,提出了区域运动相似性分割方法和相似度关联矩阵的解决方案;在运动目标检测过程中, 首先使用背景差分法提取运动区域,经过消除缺口、空洞和分离等处理,在运动区域所在范围内进行块匹配搜索和局部光流计算区域运动矢量,然后使用模糊聚类方法对运动矢量区域融合,完整的分割出粘连运动目标;在目标跟踪部分,目标跟踪建立在目标关联的基础上,提出建立连续两帧目标间距离和局部二元模式相似度关联矩阵的方法进行运动目标标定,从而实现多目标关联;使用公共视频库的图像序列进行测试,所提算法都能实现连续的跟踪和准确的运动目标分割,且处理速度快,表明了算法具有鲁棒性和适用性。     

分裂合并运动分割的多运动视觉里程计方法

目的 视觉里程计（visual odometry,VO）仅需要普通相机即可实现精度可观的自主定位,已经成为计算机视觉和机器人领域的研究热点,但是当前研究及应用大多基于场景为静态的假设,即场景中只有相机运动这一个运动模型,无法处理多个运动模型,因此本文提出一种基于分裂合并运动分割的多运动视觉里程计方法,获得场景中除相机运动外多个运动目标的运动状态。方法 基于传统的视觉里程计框架,引入多模型拟合的方法分割出动态场景中的多个运动模型,采用RANSAC（random sample consensus）方法估计出多个运动模型的运动参数实例;接着将相机运动信息以及各个运动目标的运动信息转换到统一的坐标系中,获得相机的视觉里程计结果,以及场景中各个运动目标对应各个时刻的位姿信息;最后采用局部窗口光束法平差直接对相机的姿态以及计算出来的相机相对于各个运动目标的姿态进行校正,利用相机运动模型的内点和各个时刻获得的相机相对于运动目标的运动参数,对多个运动模型的轨迹进行优化。结果 本文所构建的连续帧运动分割方法能够达到较好的分割结果,具有较好的鲁棒性,连续帧的分割精度均能达到近100%,充分保证后续估计各个运动模型参数的准确性。本文方法不仅能够有效估计出相机的位姿,还能估计出场景中存在的显著移动目标的位姿,在各个分段路径中相机自定位与移动目标的定位结果位置平均误差均小于6%。结论 本文方法能够同时分割出动态场景中的相机自身运动模型和不同运动的动态物体运动模型,进而同时估计出相机和各个动态物体的绝对运动轨迹,构建出多运动视觉里程计过程。     

Detecting slowly moving infrared targets using temporal filtering and association strategy

The special characteristics of slowly moving infrared targets, such as containing only a few pixels, shapeless edge, low signal-to-clutter ratio, and low speed, make their detection rather difficult, especially when immersed in complex backgrounds. To cope with this problem, we propose an effective infrared target detection algorithm based on temporal target detection and association strategy. First, a temporal target detection model is developed to segment the interested targets. This model contains mainly three stages, i.e., temporal filtering, temporal target fusion, and cross-product filtering. Then a graph matching model is presented to associate the targets obtained at different times. The association relies on the motion characteristics and appearance of targets, and the association operation is performed many times to form continuous trajectories which can be used to help disambiguate targets from false alarms caused by random noise or clutter. Experimental results show that the proposed method can detect slowly moving infrared targets in complex backgrounds accurately and robustly, and has superior detection performance in comparison with several recent methods.     

Cluster tracking under kinematical constraints using random matrices

Collectively moving object clusters are of particular interest in certain applications and have to be tracked as separate aggregated entities consisting of an unknown number of individuals. Group targets often will not cause as many detections since there are individual targets in the group due to limited sensor resolution capabilities. In this case, tracking and data association under the ‘one target, one detection’ assumption are no longer applicable. In this paper, ellipsoidal object extensions are modeled by random matrices, which are treated as additional state variables to be estimated. An important aspect of GMTI tracking is the incorporation of context information into the Bayesian data processing formalism. Here we consider kinematical constraints such as road maps and sensor specific characteristics such as Doppler blindness.     

Online multiple objects tracking with detection reliability prior constraint

Multi-object tracking (MOT) is one popular topic in computer vision. It remains a challenging problem in complex scenes, especially of objects with similar appearance. In this case, many existing data association strategies, which link detections among consecutive frames according appearance and motion cues, may fail to track due to unreliable detections or confused appearance and motion. To solve this problem, this paper proposed a novel online multi-object tracking method with detection reliability prior constraint. Our method integrates the trajectory estimation and detection-prediction association into a unified framework. The detection reliability prior constraint is built with the Hankel matrix from object motion model. When we build the Hankel matrix, we adaptively select a set of previous frames to predict object states and calculate the associated weights between detections and candidate objects. Data association in MOT then is estimated by maximum a posteriori (MAP) in a Bayesian framework, accompanied with both previous trajectory and the current detection reliability. Experimental results using synthetic dataset and four public challenging datasets demonstrate that, the proposed method has a good tracking performance compared with the state-of-the-art multi-object trackers.     

基于自适应时空码书的粒子滤波多目标跟踪

提出一种基于自适应时空码书检测模型的粒子滤波多目标跟踪算法。使用时空码书模型进行前景背景分割,检测出前景目标,在该模型上加入目标自适应过程。将自适应时空码书检测的结果作为粒子滤波跟踪算法的初始目标状态,通过关联算法和粒子滤波实现多目标跟踪。自适应时空码书模型能明显降低对前景目标的误检率,抑制噪声干扰。实验结果表明,该算法能够在有干扰的复杂背景下实现对运动多目标的快速捕获,并有效提高跟踪的可靠性和精度。     

交通路口视频奇异运动事件检测

提出了一种基于运动特征的交通路口视频奇异事件的检测的方法。该方法针对具体应用背景,不需要进行单个目标的跟踪和检测,而是直接运动运动特征进行检测。方法首先提取运动像素,并通过运动像素集合和运动矢量的大小判断运动目标种类（车辆或行人）。然后,在此基础上进一步分析目标的运动轨迹,将新出现的和已经有运动轨迹明显不一致的运动,和明显不符合交通规则的运动（如车辆的逆行）判定为异常运动。该方法在一定样本基础上进行了实验,实验结果表明,该方法可以较好地检测交通路口的奇异性运动事件。     

Particle Video: Long-Range Motion Estimation Using Point Trajectories

This paper describes a new approach to motion estimation in video. We represent video motion using a set of particles. Each particle is an image point sample with a long-duration trajectory and other properties. To optimize particle trajectories we measure appearance consistency along the particle trajectories and distortion between the particles. The resulting motion representation is useful for a variety of applications and cannot be directly obtained using existing methods such as optical flow or feature tracking. We demonstrate the algorithm on challenging real-world videos that include complex scene geometry, multiple types of occlusion, regions with low texture, and non-rigid deformations.