WO2019241963A1 - 基于高阶累积量的目标跟踪方法、装置及存储介质 - Google Patents

基于高阶累积量的目标跟踪方法、装置及存储介质 Download PDF

Info

Publication number
WO2019241963A1
WO2019241963A1 PCT/CN2018/092205 CN2018092205W WO2019241963A1 WO 2019241963 A1 WO2019241963 A1 WO 2019241963A1 CN 2018092205 W CN2018092205 W CN 2018092205W WO 2019241963 A1 WO2019241963 A1 WO 2019241963A1
Authority
WO
WIPO (PCT)
Prior art keywords
video frame
target
reconstruction error
occluded
subspace
Prior art date
Application number
PCT/CN2018/092205
Other languages
English (en)
French (fr)
Inventor
李良群
谢维信
刘宗香
Original Assignee
深圳大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 深圳大学 filed Critical 深圳大学
Priority to PCT/CN2018/092205 priority Critical patent/WO2019241963A1/zh
Publication of WO2019241963A1 publication Critical patent/WO2019241963A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments

Definitions

  • the invention relates to the field of target tracking, in particular to a method, device and storage medium for target tracking based on high-order cumulants.
  • Online target tracking is a hot research topic in computer vision, which is of great significance for high-level visual research such as motion recognition, behavior analysis, and scene understanding, and has a wide range of applications in video surveillance, intelligent robots, and human-computer interaction. prospect.
  • Occlusion is one of the most difficult problems. Occlusion makes part or the whole of the tracking target invisible, causing the loss of target information, and the duration of the occlusion is unpredictable.
  • the occlusion of the tracking target may be caused by other moving objects in the video, static objects in the background, or the target itself. In occlusion situations, tracking drift may occur.
  • Many algorithms in the prior art can process occlusion to improve the accuracy of target tracking, but the efficiency of occlusion detection is relatively low, and it is often difficult to determine that the target is occluded when the target has been occluded for many frames, which is not conducive to occlusion processing.
  • the present invention proposes a target tracking method based on a high-order cumulant.
  • the method includes: using a first high-order cumulant to determine whether an object in the k-th video frame is occluded.
  • the cumulant corresponds to the first reconstruction error of the target image in the k-th video frame in the subspace; using the motion model and the target state information in the k-th video frame to predict and extract multiple predicted particles, the k-th video frame is
  • the motion model of the occluded and unoccluded targets is different; the reconstruction error of the predicted image block corresponding to each predicted particle in the k + 1th video frame in the subspace is calculated separately; each reconstruction corresponding prediction is calculated using the reconstruction error
  • the importance weight of the particles use the second high-order cumulant to determine whether the occluded object in the k-th video frame is still occluded in the k + 1 video frame.
  • the second high-order cumulant corresponds to the likelihood image block in The second reconstruction error of the subspace, the likelihood image block is the predicted image block corresponding to the predicted particle with the greatest importance weight; if the target is blocked in the k-th video frame and the k + 1-th video frame, the trajectory is predicted Value as the target image in the k + 1 video frame, otherwise the likelihood image block is used as the target image in the k + 1 video frame; use the target image in the k + 1 video frame to obtain the k + 1 video frame Target status information.
  • the present invention provides a target tracking device based on a high-order cumulant.
  • the device includes at least one processor, which works alone or in cooperation.
  • the processor is configured to execute instructions to implement the foregoing method.
  • the present invention provides a readable storage medium that stores instructions, and implements the foregoing methods when the instructions are executed.
  • the beneficial effect of the present invention is: by using the first high-order cumulant, it is judged whether the target in the k-th video frame is occluded, and different motion models and target state information in the k-th video frame are used for the blocked and unoccluded targets Predict and extract multiple predicted particles; calculate the reconstruction error of each predicted particle corresponding to the predicted image block in the k + 1 video frame in the subspace separately; use the reconstruction error to calculate the importance of each predicted particle sexual weight; use the second high-order cumulant to determine whether the occluded object in the k-th video frame is still occluded in the k + 1 video frame, and the second high-order cumulant corresponds to the likelihood of the image block in the subspace Second reconstruction error, the likelihood image block is the predicted image block corresponding to the predicted particle with the greatest importance weight; if the target is blocked in both the kth video frame and the k + 1th video frame, the trajectory prediction value is used as the first the target image in the k + 1 video frame, otherwise
  • FIG. 1 is a schematic flowchart of a first embodiment of a target tracking method based on a high-order cumulant according to the present invention
  • FIG. 2 is a schematic diagram of determining an occlusion state based on a third-order cumulant of reconstruction errors in a specific embodiment of the present invention
  • FIG. 3 is a key frame corresponding to the third-order cumulant curve in FIG. 2;
  • FIG. 4 is a schematic flowchart of a second embodiment of a target tracking method based on a high-order cumulant according to the present invention
  • FIG. 5 is a schematic structural diagram of a first embodiment of a target tracking device based on a high-order cumulant of the present invention
  • FIG. 6 is a schematic structural diagram of a first embodiment of a readable storage medium according to the present invention.
  • a first embodiment of a method for tracking a target based on a high-order cumulant of the present invention includes:
  • S1 Use the first high-order cumulant to determine whether an object in the k-th video frame is blocked.
  • the first higher-order cumulant corresponds to the first reconstruction error of the target image in the k-th video frame in the subspace, and k is a positive integer.
  • the problem of target tracking can be regarded as the problem of dynamic estimation of target state in video frames.
  • Z 1: k + 1 ) satisfy:
  • X k ) is the target's motion model and is used to describe the change of the target's motion between two consecutive frames;
  • X k + 1 ) is the target's observation model, Represents the observed likelihood function.
  • Principal component analysis can be used to establish the observation model of the target.
  • the observation samples can include target images and predicted image blocks.
  • the subspace U [u 1 , u 2 , ..., u u ] includes u orthogonal orthogonal uncorrelated base vectors, the target image Z k and the sample mean Z in the k- th video frame.
  • the difference can be expressed by the linear form of the base vector in the subspace U as:
  • the reconstruction error can be considered as a small variance Gaussian white noise. If the tracking target is occluded, the reconstruction error will become the sum of white Gaussian noise and random signals, showing a non-Gaussian nature. Therefore, the feature that the high-order cumulant (greater than the second order) of the Gaussian signal is all zero can be used to detect the occlusion situation of the target.
  • s (j) is a non-Gaussian random signal caused by tracking target occlusion
  • v (j) is Gaussian noise
  • s (j) and v (j) are independent of each other.
  • a first order cumulant reconstruction error e k is defined as:
  • T greater than 0 is selected as a preset threshold and compared with the first high-order cumulant C k .
  • the first higher-order cumulant C k can be obtained as:
  • the first high-order accumulation amount C k is greater than the preset threshold T, it is determined that the target in the k-th video frame is blocked, otherwise it is determined that the target in the k-th video frame is not blocked.
  • FIG. 2 is a schematic diagram of occlusion state determination based on a third-order cumulant of reconstruction error.
  • the third-order cumulant curve of reconstruction error is at a horizontal threshold Above, it means that the target is occluded, otherwise it means that the target is not occluded.
  • the key frames corresponding to the third-order cumulant curve in Figure 2 are shown in Figure 3.
  • Target state information X k ⁇ x k , y k , s k , ⁇ k ⁇ in the k-th video frame, where x k and y k are the x-coordinate and y of the center point position of the target image in the k-th video frame.
  • the coordinates, sk and ⁇ k are the ratio of the target image to the standard size and the aspect ratio in the k-th video frame, respectively.
  • the motion model of the occluded and unoccluded targets in the k-th video frame is different.
  • a random walk model is used for prediction; otherwise, a second-order autoregressive model is used.
  • X k ) of the random walk model is:
  • X k + 1 is the target state information in the k + 1th video frame
  • N () is the normal distribution
  • ⁇ is the diagonalized covariance matrix
  • the diagonal elements of ⁇ are ⁇ x represents the standard deviation of the center point position x coordinate, ⁇ y represents the standard deviation of the center point position y coordinate, ⁇ s represents the standard deviation of the ratio to the standard size, and ⁇ ⁇ represents the standard deviation of the aspect ratio to the standard size.
  • the second-order autoregressive model assumes that the difference between X k + 1 and X k is similar to the difference between X k and X k-1 , specifically:
  • W k + 1 is Gaussian white noise.
  • the extracted N predicted particles form a predicted particle set
  • the i-th predicted particle The corresponding predicted image block in the k + 1th video frame Reconstruction error in subspace U for:
  • Reconstruction error The gray value of the j-th pixel in the Reconstruction error The number of pixels in the pixel.
  • ⁇ represents the standard deviation of the standard normal distribution
  • is the second norm
  • the second higher-order cumulant C k + 1, max corresponds to the second reconstruction error e k + 1, max of the likelihood image block Z k + 1, max in the subspace U.
  • the likelihood image block is an importance weight The predicted image block corresponding to the largest predicted particle.
  • the second reconstruction error e k + 1, max is:
  • f k + 1, max (j) is the gray value of the j-th pixel point in the second reconstruction error
  • M k + 1, max is the number of pixel points in the second reconstruction error
  • the calculation of the second reconstruction error e k + 1, max has actually been completed in S3. After S4 is completed, the corresponding likelihood image block Z k + 1 can be determined by finding the predicted particle with the largest importance weight . max and the second reconstruction error e k + 1, max .
  • the second high-order accumulation amount C k + 1, max is greater than the preset threshold T, it is determined that the target in the k + 1 video frame is blocked, otherwise it is determined that the target in the k + 1 video frame is not blocked.
  • the predicted particle with the highest importance weight has the greatest similarity to the target template (ie, the sample mean). If the target is not blocked in the k-th video frame or the k + 1-th video frame, the likelihood can be directly based on the principle of maximum similarity.
  • the image block serves as the target image in the k + 1th video frame. If both the k-th video frame and the k + 1-th video frame are occluded, the similarity between the predicted particle and the target template cannot be used to locate the target, so the trajectory prediction value, that is, the calculation result of formula (4) is used as the k + th The target image in 1 video frame.
  • the first high-order cumulant is used to determine whether the target in the k-th video frame is occluded.
  • different motion models and target state information in the k-th video frame are used to predict.
  • the likelihood image block is the predicted image block corresponding to the predicted particle with the largest importance weight; if the target is blocked in the k-th video frame and the k + 1-th video frame, the trajectory prediction value is used as the k-th The target image in the +1 video frame, otherwise use the likelihood image block as the target image in the k + 1 video frame; use the target image in the k + 1 video frame to obtain the target state information in the k + 1 video frame .
  • the second embodiment of the target tracking method based on the high-order cumulant of the present invention is based on the first embodiment of the target tracking method of the present invention based on the high-order cumulant.
  • S6 it further includes:
  • the appearance of the target and the background will change continuously, and the appearance of the target should change accordingly. Updating the subspace and sample mean in time can ensure the effectiveness and accuracy of tracking.
  • the central data matrix of training image set A is Central data matrix
  • the singular value decomposition obtains the subspace U and the feature vector ⁇ .
  • the new image set B includes target images in k + 1 video frames, and m and n are integers greater than or equal to 1.
  • the joint matrix R is:
  • a forgetting factor f is set when the sample mean is updated. Updated sample mean at this point for:
  • the target image cannot accurately reflect the target information.
  • the target template is used to update the target template, the error of the target template will be increased, so the target will not be blocked if the target is blocked in the k + 1th video frame. Perform this step.
  • the subspace and sample mean may not be updated. For example, before this step, it may be judged whether the update conditions are met, and the update is performed if it is met, otherwise the update is not performed.
  • the update condition may include that the number of consecutive video frames whose target is not occluded is greater than a threshold, and the like.
  • HOCPT high-order cumulant-based particle filter pedestrian target tracking algorithm
  • the first evaluation criterion is defined as the Euclidean distance between the center position of the tracking target and the accurate position manually calibrated.
  • the second evaluation criterion is defined as the ratio of the intersection and union of the target tracking rectangle and the target true rectangle.
  • the third evaluation criterion, the success rate based on the overlap mechanism is defined as the ratio of the number of frames successfully tracked to the target in the entire video sequence to the total number of frames in the entire video sequence.
  • the criterion for judging whether the target is successfully tracked is whether the overlap rate is greater than a given threshold T 0 , and the threshold T 0 is generally set to 0.5.
  • the HOCPT algorithm is compared with the IVT algorithm, the TLD (Tracking-Learning-Detection) algorithm and the VTD (Visual Tracking Decomposition) algorithm.
  • the five test sequences all include occlusion cases.
  • the sequence Walking includes short-time local occlusion and scale transformation.
  • the scale change prevents the TLD algorithm from accurately detecting the target, which leads to tracking offset.
  • the HOCPT algorithm, VTD algorithm, and IVT algorithm can accurately track the target. Due to local occlusion and similar interference in the sequence Walking2, the TLD and VTD algorithms all follow the wrong target, while the HOCPT algorithm and the IVT algorithm both have a good tracking effect.
  • the target is completely occluded for a short time, a large tracking offset occurs between the VTD algorithm and the IVT algorithm, and the HOCPT algorithm and the TLD algorithm both achieve good tracking results.
  • the HOCPT algorithm and the TLD algorithm are completely completed at the target When the occlusion reappears, the target can be quickly and accurately captured. Moreover, the HOCPT algorithm scales adaptively, which makes the average center position error of the tracking result of the HOCPT algorithm smaller, the average overlap rate larger, and the tracking result more stable. For the test sequence Woman, the target is partially occluded for a long time and there are complex background changes, making the four algorithms HOCPT, TLD, VTD, and IVT fail to achieve good tracking results.
  • the tracking success rate, average center position error, and average overlap rate of different algorithms on each test sequence are shown in Table 1, Table 2, and Table 3, respectively.
  • the HOCPT algorithm has a higher success rate, a smaller center position error, and a higher overlap rate, and generally has stronger robustness and stability.
  • a first embodiment of a target tracking device based on a high-order cumulant of the present invention includes a processor 110. Only one processor 110 is shown in the figure, and the actual number can be more. The processors 110 may work individually or in cooperation.
  • the processor 110 controls the operation of the target tracking device based on the high-order cumulant.
  • the processor 110 may also be referred to as a CPU (Central Processing Unit).
  • the processor 110 may be an integrated circuit chip and has a processing capability of a signal sequence.
  • the processor 110 may also be a general-purpose processor, a digital signal sequence processor (DSP), an application specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA) or other programmable logic device, a discrete gate or transistor logic device, and discrete hardware Components.
  • DSP digital signal sequence processor
  • ASIC application specific integrated circuit
  • FPGA off-the-shelf programmable gate array
  • a general-purpose processor may be a microprocessor or the processor may be any conventional processor or the like.
  • the processor 110 is configured to execute instructions to implement the method provided by the first or second embodiment of the target tracking method based on the high-order cumulant of the present invention.
  • the first embodiment of the readable storage medium of the present invention includes a memory 210 that stores instructions that, when executed, implement the first or second embodiment of the target tracking method of the present invention based on a high-order cumulant. The method provided.
  • the memory 210 may include a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a flash memory (Flash), a hard disk, an optical disk, and the like.
  • ROM read-only memory
  • RAM random access memory
  • flash flash memory
  • the disclosed methods and devices may be implemented in other ways.
  • the device implementation described above is only schematic.
  • the division of the modules or units is only a logical function division.
  • multiple units or components may The combination can either be integrated into another system, or some features can be ignored or not implemented.
  • the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, which may be electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, may be located in one place, or may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objective of the solution of this embodiment.
  • each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may be separately physically included, or two or more units may be integrated into one unit.
  • the above integrated unit may be implemented in the form of hardware or in the form of software functional unit.
  • the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it may be stored in a computer-readable storage medium.
  • the technical solution of the present invention essentially or part that contributes to the existing technology or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium , Including a number of instructions for causing a computer device (which may be a personal computer, a server, or a network device) or a processor to perform all or part of the steps of the method described in each embodiment of the present invention.
  • the aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disks or optical disks and other media that can store program codes .

Landscapes

  • Engineering & Computer Science (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)

Abstract

一种基于高阶累积量的目标跟踪方法及装置，该方法包括：利用第一高阶累积量判断第k视频帧中的目标是否被遮挡(S1)；利用运动模型及第k视频帧中的目标状态信息预测并抽取得到多个预测粒子(S2)；分别计算每个预测粒子在第k+1视频帧中对应的预测图像块在子空间的重构误差(S3)及对应的每个预测粒子的重要性权值(S4)；利用第二高阶累积量判断第k视频帧中被遮挡的目标在第k+1视频帧中是否仍被遮挡(S5)；若是则将轨迹预测值否则将似然图像块作为第k+1视频帧中的目标图像(S6)；利用第k+1视频帧中的目标图像获取第k+1视频帧中的目标状态信息(S7)。

Description

基于高阶累积量的目标跟踪方法、装置及存储介质 【技术领域】
本发明涉及目标跟踪领域，特别是涉及一种基于高阶累积量的目标跟踪方法、装置及存储介质。
【背景技术】
在线目标跟踪是计算机视觉中的一个热点研究课题，其对于动作识别、行为分析、场景理解等高层次的视觉研究具有重要意义，并且在视频监控、智能机器人、人机交互等领域有着广泛的应用前景。
在目标跟踪的许多挑战中，遮挡是最棘手的问题之一。遮挡使跟踪目标的部分或者整体不可见，造成目标信息丢失，并且遮挡持续的时间长短不可预知。跟踪目标被遮挡可能是由视频中其它运动物、背景中静止的物体或者目标本身引起的。在遮挡情形下，可能会出现跟踪漂移。现有技术中的很多算法可以对遮挡进行处理以提高目标跟踪的准确度，但是遮挡检测的效率比较低，往往是在目标已经被遮挡了很多帧才能判断出目标被遮挡，不利于遮挡处理。
【发明内容】
为了至少部分解决以上问题，本发明提出了一种基于高阶累积量的目标跟踪方法，该方法包括：利用第一高阶累积量判断第k视频帧中的目标是否被遮挡，第一高阶累积量对应于第k视频帧中的目标图像在子空间的第一重构误差；利用运动模型及第k视频帧中的目标状态信息预测并抽取得到多个预测粒子，第k视频帧中被遮挡与未被遮挡的目标的运动模型不同；分别计算每个预测粒子在第k+1视频帧中对应的预测图像块在子空间的重构误差；分别利用重构误差计算对应的每个预测粒子的重要性权值；利用第二高阶累积量判断第k视频帧中被遮挡的目标在第k+1视频帧中是否仍被遮挡，第二高阶累积量对应于似然图像块在子空间的第二重构误差，似然图像块是重要性权值最大的预测粒子对应的预测图像块；若第k视频帧和第k+1视频帧中目标均被遮挡，则将轨迹预测值作为第k+1视频帧中的目标图像，否则将似然图像块作为第k+1视频帧中的目标图像；利用第k+1视频帧中的目标图像获取第k+1视频帧中的目标状 态信息。
为了解决上述技术问题，本发明提供了一种基于高阶累积量的目标跟踪装置，该装置包括至少一个处理器，单独或协同工作，处理器用于执行指令以实现前述的方法。
为了解决上述技术问题，本发明提供了一种可读存储介质，存储有指令，指令被执行时实现前述的方法。
本发明的有益效果是：通过利用第一高阶累积量判断第k视频帧中的目标是否被遮挡，对于遮挡和未被遮挡的目标使用不同的运动模型及第k视频帧中的目标状态信息预测并抽取得到多个预测粒子；分别计算每个预测粒子在第k+1视频帧中对应的预测图像块在子空间的重构误差；分别利用重构误差计算对应的每个预测粒子的重要性权值；利用第二高阶累积量判断第k视频帧中被遮挡的目标在第k+1视频帧中是否仍被遮挡，第二高阶累积量对应于似然图像块在子空间的第二重构误差，似然图像块是重要性权值最大的预测粒子对应的预测图像块；若第k视频帧和第k+1视频帧中目标均被遮挡，则将轨迹预测值作为第k+1视频帧中的目标图像，否则将似然图像块作为第k+1视频帧中的目标图像；利用第k+1视频帧中的目标图像获取第k+1视频帧中的目标状态信息。在目标未被遮挡的情况下，其重构误差是一个小方差的高斯白噪声，在目标被遮挡的情况下，其重构误差是噪声与随机信号之和。利用高阶累积量对于高斯噪声良好的抑制作用，可以快速检测到被遮挡时重构误差中的随机信号，从而能够及时并正确地判断遮挡的存在并进行遮挡处理，提高了跟踪的准确度。
【附图说明】
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1是本发明基于高阶累积量的目标跟踪方法第一实施例的流程示意图；
图2是本发明一具体实施例中基于重构误差的三阶累积量进行遮挡状态判断的示意图；
图3是图2中三阶累积量曲线对应的关键帧；
图4是本发明基于高阶累积量的目标跟踪方法第二实施例的流程示意图；
图5是本发明基于高阶累积量的目标跟踪装置第一实施例的结构示意图；
图6是本发明可读存储介质第一实施例的结构示意图。
【具体实施方式】
下面结合附图和实施例对本发明进行详细说明。以下实施例中不冲突的可以相互结合。显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。
本发明的说明书和权利要求书及上述附图中的术语“第一”、“第二”、“第三”“第四”等(如果存在)是用于区别类似的对象，而不必用于描述特定的顺序或先后次序。应该理解这样使用的数据在适当情况下可以互换，以便这里描述的本发明的实施例例如能够以除了在这里图示或描述的那些以外的顺序实施。此外，术语“包括”和“具有”以及他们的任何变形，意图在于覆盖不排他的包含，例如，包含了一系列步骤或单元的过程、方法、系统、产品或设备不必限于清楚地列出的那些步骤或单元，而是可包括没有清楚地列出的或对于这些过程、方法、产品或设备固有的其它步骤或单元。
如图1所示，本发明基于高阶累积量的目标跟踪方法第一实施例包括：
S1：利用第一高阶累积量判断第k视频帧中的目标是否被遮挡。
第一高阶累积量对应于第k视频帧中的目标图像在子空间的第一重构误差，k为正整数。
目标跟踪问题可视为视频帧中目标状态动态估计问题。根据贝叶斯定理，给定一组观测序列Z 1:k＝{Z 1,Z 2,…,Z k+1}，k+1时刻行人目标状态的后验概率p(X k+1|Z 1:k+1)满足：
p(X k+1|Z 1:k+1)∝p(Z k+1|X k+1)∫p(X k+1|X k)p(X k|Z 1:k)dX k    (15)
其中，p(X k+1|X k)为目标的运动模型，用于描述两个连续帧之间目标的运动变化；p(Z k+1|X k+1)为目标的观测模型，表示观测似然函数。
可以使用主成分分析的方法建立目标的观测模型。通过对训练样本进行主成分分析得到一组子空间的基向量，然后利用子空间计算观测样本的重构误差，观测样本可以包括目标图像及预测图像块。
具体的，子空间U＝[u 1,u 2,...,u u]包括u个两两之间正交不相关的基向量，第k视频帧中的目标图像Z k与样本均值Z之差可以用子空间U中的基向量线型表示为：
Figure PCTCN2018092205-appb-000001
其中e k为第一重构误差，W＝(w 1,w 2,...,w u) T为
Figure PCTCN2018092205-appb-000002
在子空间U上的投影，即
Figure PCTCN2018092205-appb-000003
代入式(16)可得，第一重构误差ek为：
Figure PCTCN2018092205-appb-000004
第一重构误差e k可以用序列{f k(j),j＝0,1,...M k-1}来表示，f k(j)为第一重构误差中第j个像素点的灰度值，M k为第一重构误差中像素点的个数。
若目标处于正常运动状态，即未被遮挡，重构误差可以认为是一个小方差的高斯白噪声。若跟踪目标被遮挡，重构误差会变为高斯白噪声和随机信号之和，呈现出非高斯特性。因此，可以利用高斯信号的高阶累积量(大于二阶)均为零的特点来实现对目标遮挡情况的检测。
对于第一重构误差e k＝{f k(j),j＝0,1,...M k-1}，H 0为正常状态，H 1为遮挡状态，构造二元假设检验：
Figure PCTCN2018092205-appb-000005
其中，s(j)是由跟踪目标遮挡引起的非高斯随机信号，v(j)为高斯噪声，且s(j)与v(j)相互独立。
设第一重构误差e k均值为0，根据零均值随机过程的三阶累积量等于其三阶矩，第一重构误差e k的三阶累积量定义为：
Figure PCTCN2018092205-appb-000006
其中E()表示数学期望。v(j)是高斯噪声，所以
Figure PCTCN2018092205-appb-000007
则有
Figure PCTCN2018092205-appb-000008
其中C 3s(g,h)和C 3v(g,h)分别为随机信号和高斯噪声的三阶累积量，由于|C 3s(g,h)|在原点取峰值，即：|C 3s(g,h)|≤|C 3s(0,0)|，可知
Figure PCTCN2018092205-appb-000009
因此：
Figure PCTCN2018092205-appb-000010
在实际应用中，数据长度有限，高斯噪声的三阶累积量估计并不为零，因此选取大于0的T作为预设阈值与第一高阶累积量C k进行比较。
根据式(18)，可得第一高阶累积量C k为：
Figure PCTCN2018092205-appb-000011
若第一高阶累积量C k大于预设阈值T，则判定第k视频帧中的目标被遮挡，否则判定第k视频帧中的目标未被遮挡。
举例说明，在本发明一具体实施例中，图2为基于重构误差的三阶累积量进行遮挡状态判断的示意图，在图2中，如果重构误差的三阶累积量曲线在水平的阈值之上，则意味着目标被遮挡，否则意味着目标未被遮挡。图2中三阶累积量曲线对应的关键帧如图3所示。
结合图2和图3可以看出：在65帧时目标刚刚进入遮挡状态，对应的重构误差的三阶累积量大于阈值，判为遮挡状态；在86帧时目标重新出现，对应的重构误差的三阶累积量小于阈值，判为未遮挡的运动状态，并且在目标被遮挡的时间段内，其对应的重构误差图像的三阶累积量大于阈值。可知基于重构误差的三阶累积量的遮挡检测方法可以准确地判断目标进入与离开遮挡的时刻。
S2：利用运动模型及第k视频帧中的目标状态信息预测并抽取得到多个预测粒子。
第k视频帧中的目标状态信息X k＝{x k,y k,s k,α k}，其中x k和y k分别为第k视频帧中的目标图像中心点位置的x坐标和y坐标，s k和α k分别为第k视频帧中的目标图像与标准大小的比例以及宽高比。
第k视频帧中被遮挡与未被遮挡的目标的运动模型不同。
可选的，若第k视频帧中目标被遮挡，则利用随机游走模型来进行预测；否则利用二阶自回归模型。
随机游走模型的状态转移概率p(X k+1|X k)为：
p(X k+1|X k)＝N(X k+1|X k,Ψ)         (3)
其中，X k+1为第k+1视频帧中的目标状态信息，N()为正态分布，ψ为对角化协方差矩阵，ψ的对角元素分别为
Figure PCTCN2018092205-appb-000012
δ x表示中心点位置x坐标的标准差，δ y表示中心点位置y坐标的标准差，δ s表示与标准大小的比例的标准差，δ α表示与标准大小的宽高比的标准差。
二阶自回归模型假设X k+1和X k之间的差异与X k和X k-1之间的差异相近，具体为：
X k+1-X k＝X k-X k-1+W k+1          (4)
其中，W k+1为高斯白噪声。
抽取的N个预测粒子组成预测粒子集
Figure PCTCN2018092205-appb-000013
S3：分别计算每个预测粒子在第k+1视频帧中对应的预测图像块在子空间的重构误差。
参考第一重构误差e k的计算过程，第i个预测粒子
Figure PCTCN2018092205-appb-000014
在第k+1视频帧中对应的预测图像块
Figure PCTCN2018092205-appb-000015
在子空间U的重构误差
Figure PCTCN2018092205-appb-000016
为：
Figure PCTCN2018092205-appb-000017
其中，
Figure PCTCN2018092205-appb-000018
为重构误差
Figure PCTCN2018092205-appb-000019
中第j个像素点的灰度值，
Figure PCTCN2018092205-appb-000020
为重构误差
Figure PCTCN2018092205-appb-000021
中像素点的个数。
S4：分别利用重构误差计算对应的每个预测粒子的重要性权值。
第i个预测粒子的重要性权值
Figure PCTCN2018092205-appb-000022
满足
Figure PCTCN2018092205-appb-000023
为第i个预测粒子
Figure PCTCN2018092205-appb-000024
的观测似然函数，具体为：
Figure PCTCN2018092205-appb-000025
其中，δ表示标准正态分布的标准差，||·||为二范数。
S5：利用第二高阶累积量判断第k视频帧中被遮挡的目标在第k+1视频帧中是否仍被遮挡。
第二高阶累积量C k+1,max对应于似然图像块Z k+1,max在子空间U的第二重构误差e k+1,max，似然图像块是重要性权值最大的预测粒子对应的预测图像块。
第二重构误差e k+1,max为：
Figure PCTCN2018092205-appb-000026
其中，f k+1,max(j)为第二重构误差中第j个像素点的灰度值，M k+1,max为第二重构误差中像素点的个数。
第二重构误差e k+1,max的计算在S3中实际上已经完成，在S4完成后，找到重要性权值最大的预测粒子即可确定其对应的似然图像块Z k+1,max以及第二重构误差e k+1,max。
参考第一高阶累积量C k的计算过程，第二高阶累积量C k+1,max为：
Figure PCTCN2018092205-appb-000027
若第二高阶累积量C k+1,max大于预设阈值T，则判定第k+1视频帧中的目标被遮挡，否则判定第k+1视频帧中的目标未被遮挡。
为了节省计算量，可以选择仅为在第k视频帧中被遮挡的目标执行本步骤，即只在目标在第k视频帧中被遮挡的情况下，利用第二重构误差计算第二高阶累积量之后与预设阈值比较判断。
S6：若第k视频帧和第k+1视频帧中目标均被遮挡，则将轨迹预测值作为第k+1视频帧中的目标图像，否则将似然图像块作为第k+1视频帧中的目标图像。
重要性权值最大的预测粒子与目标模板(即样本均值)相似度最大，若第k视频帧或第k+1视频帧中目标未被遮挡，可以基于相似度最大的原则，直接将似然图像块作为第k+1视频帧中的目标图像。若第k视频帧和第k+1视频帧中目标均被遮挡，预测粒子与目标模板的相似度无法用来定位目标，因此将轨迹预测值，即式(4)的计算结果作为第k+1视频帧中的目标图像。
S7：利用第k+1视频帧中的目标图像获取第k+1视频帧中的目标状态信息。
根据第k+1视频帧中的目标图像的位置信息计算可以得到k+1视频帧中的目标状态信息X k+1＝{x k+1,y k+1,s k+1,α k+1}。
通过本实施例的实施，利用第一高阶累积量判断第k视频帧中的目标是否被遮挡，对于遮挡和未被遮挡的目标使用不同的运动模型及第k视频帧中的目标状态信息预测并抽取得到多个预测粒子；分别计算每个预测粒子在第k+1视频帧中对应的预测图像块在子空间的重构误差；分别利用重构误差计算对应的每个预测粒子的重要性权值；利用第二高阶累积量判断第k视频帧中被遮挡的目标在第k+1视频帧中是否仍被遮挡，第二高阶累积量对应于似然图像块在子空间的第二重构误差，似然图像块是重要性权值最大的预测粒子对应的预测图像块；若第k视频帧和第k+1视频帧中目标均被遮挡，则将轨迹预测值作为第k+1视频帧中的目标图像，否则将似然图像块作为第k+1视频帧中的目标图像；利用第k+1视频帧中的目标图像获取第k+1视频帧中的目标状态信息。利用高阶累积量对于高斯噪声良好的抑制作用，可以快速检测到被遮挡时重构误差中的随机信号，从而能够及时并正确地判断遮挡的存在并进行遮挡处理，提高了跟踪的准确度。
如图4所示，本发明基于高阶累积量的目标跟踪方法第二实施例，是在本 发明基于高阶累积量的目标跟踪方法第一实施例的基础上，S6之后进一步包括:
S8:目标在第k+1视频帧中未被遮挡的情况下，至少利用第k+1视频帧中的目标图像增量更新子空间和样本均值。
目标跟踪过程中，目标和背景的外观会不断地发生变化，相应地目标外观表示也应该随之发生变化。及时更新子空间和样本均值可以保证跟踪的有效性和准确性。
具体的，更新前的子空间U对应的训练图像集为A＝{Z 1，Z 2，…，Z n}，n为训练图像集A中图像的数量，训练图像集A的均值为样本均值
Figure PCTCN2018092205-appb-000028
训练图像集A的中心数据矩阵为
Figure PCTCN2018092205-appb-000029
中心数据矩阵
Figure PCTCN2018092205-appb-000030
奇异值分解得到子空间U和特征向量Σ，新增图像集为B＝{Z n+1,Z n+2,…,Z n+m}，m为新增图像集B中图像的数量，新增图像集B包括k+1视频帧中的目标图像，m和n为大于或等于1的整数。
更新之后的全部训练样本为C＝{A,B}＝{Z 1,Z 2,…,Z n+m}。
更新后的子空间U'为：
Figure PCTCN2018092205-appb-000031
其中
Figure PCTCN2018092205-appb-000032
是对联合矩阵R进行奇异值分解得到的：
Figure PCTCN2018092205-appb-000033
联合矩阵R为：
Figure PCTCN2018092205-appb-000034
Figure PCTCN2018092205-appb-000035
其中Orth()为执行正交化；
Figure PCTCN2018092205-appb-000036
其中
Figure PCTCN2018092205-appb-000037
为新增图像集B的均值；
在目标跟踪中，跟踪一个外观有变化的目标时，往往希望最新捕获的目标图像占较大的比重而先前的目标图像占较小的比重，最新的目标图像相较于先前的目标图像可以更好地用于表示目标外观。为了平衡新旧目标图像对目标外观的影响，在更新样本均值时设置遗忘因子f。此时更新后的样本均值
Figure PCTCN2018092205-appb-000038
为：
Figure PCTCN2018092205-appb-000039
目标被遮挡时，目标图像并不能准确的反映目标信息，此时如果用目标图像来更新目标模板，会增加目标模板的误差，因此目标在第k+1视频帧中被遮挡的情况下不会执行本步骤。
在其他实施例中，目标在第k+1视频帧中未被遮挡的情况下也可以不更新子空间和样本均值。例如，本步骤之前可以先判断是否满足更新条件，若满足才进行更新，否则不更新。更新条件可以包括目标未被遮挡的连续视频帧数量大于一阈值等。
下面为对本发明基于高阶累积量的目标跟踪方法第二实施例提出的基于高阶累积量的粒子滤波行人目标跟踪算法(HOCPT)进行实验验证的结果。
本次实验在Windows 7系统下进行，硬件设备参数为英特尔Core(TM)i7-4790，主频为3.60GHZ，4G内存。实验代码的编写与调试以及实验数据实验结果图像的采集采用Matlab R2014a环境。为了验证算法对行人目标跟踪的有效性，在测试数据集中选择了五个包含行人目标的视频(a)Walking，(b)Walking2，(c)Jogging1，(d)Jogging2，(e)Woman作为测试序列进行实验。
采用中心位置误差(Center Location Error,CLE)，重叠率(Overlap Rate,OR)，以及基于重叠机制的成功率(Success Rate,SR)三种评价机制作为定量分析的标准。第一种评价标准——中心位置误差的定义为跟踪目标的中心位置和手工标定的准确位置之间的欧式距离。第二种评价标准——重叠率的定义为目标跟踪矩形区域与目标真实矩形区域的交集与并集的比值。第三种评价标准——基于重叠机制的成功率的定义为在整个视频序列中成功跟踪到目标的帧数占整个视频序列总帧数的比值。判断是否成功跟踪到目标的标准为重叠率是否大于给定的阈值T 0,阈值T 0一般设置为0.5。
为了分析算法性能，将HOCPT算法与IVT算法、TLD(Tracking-Learning-Detection,跟踪-学习-检测)算法和VTD(Visual Tracking Decomposition,视觉跟踪分解)算法进行对比。
5个测试序列均包括遮挡情形。序列Walking中包含有短时间局部遮挡以及尺度变换，尺度变化使得TLD算法无法准确检测到目标，从而导致跟踪偏移，而HOCPT算法、VTD算法与IVT算法均能准确跟踪到目标。序列Walking2中由于存在局部遮挡以及相似物干扰，TLD，VTD算法均跟错目标，而HOCPT算法与IVT算法均有较好的跟踪效果。序列Jogging1，Jogging2中目标被短时间完全遮挡，VTD算法与IVT算法发生了较大的跟踪偏移，而HOCPT算法以 及TLD算法均取得了较好的跟踪效果，HOCPT算法与TLD算法在目标被完全遮挡又重新出现时均能快速准确地捕获到目标，并且，由于HOCPT算法尺度自适应，使得HOCPT算法跟踪结果平均中心位置误差更小，平均重叠率更大，跟踪结果更稳定。对于测试序列Woman，目标被长时间局部遮挡以及存在复杂的背景变化，使得HOCPT算法，TLD，VTD，IVT四种算法均未能取得较好的跟踪结果。
不同算法在各测试序列上的跟踪成功率、平均中心位置误差、平均重叠率分别如表1,表2,表3所示。
Figure PCTCN2018092205-appb-000040
表1
Figure PCTCN2018092205-appb-000041
表2
Figure PCTCN2018092205-appb-000042
表3
结合上述表格可知，HOCPT算法具有较高的成功率、较小的中心位置误差以及较高的重叠率，总体上具有较强的鲁棒性以及稳定性。
如图5所示，本发明基于高阶累积量的目标跟踪装置第一实施例包括：处理器110。图中只画出了一个处理器110，实际数量可以更多。处理器110可以单独或者协同工作。
处理器110控制基于高阶累积量的目标跟踪装置的操作，处理器110还可以称为CPU(Central Processing Unit，中央处理单元)。处理器110可能是一种集成电路芯片，具有信号序列的处理能力。处理器110还可以是通用处理器、数字信号序列处理器(DSP)、专用集成电路(ASIC)、现成可编程门阵列(FPGA) 或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
处理器110用于执行指令以实现本发明基于高阶累积量的目标跟踪方法第一或第二实施例所提供的方法。
如图6所示，本发明可读存储介质第一实施例包括存储器210，存储器210存储有指令，该指令被执行时实现本发明基于高阶累积量的目标跟踪方法第一或第二实施例所提供的方法。
存储器210可以包括只读存储器(ROM，Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、闪存(Flash Memory)、硬盘、光盘等。
在本发明所提供的几个实施例中，应该理解到，所揭露的方法和装置，可以通过其它的方式实现。例如，以上所描述的装置实施方式仅仅是示意性的，例如，所述模块或单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或单元的间接耦合或通信连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施方式方案的目的。
另外，在本发明各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理包括，也可以两个或两个以上单元集成在一个单元中。上述集成的单元既可以采用硬件的形式实现，也可以采用软件功能单元的形式实现。
所述集成的单元如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本发明的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的全部或部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)或处理器(processor)执行本发明各个实施方式所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(ROM， Read-Only Memory)、随机存取存储器(RAM，Random Access Memory)、磁碟或者光盘等各种可以存储程序代码的介质。
以上所述仅为本发明的实施方式，并非因此限制本发明的专利范围，凡是利用本发明说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本发明的专利保护范围内。

Claims (12)

  1. 一种基于高阶累积量的目标跟踪方法，其特征在于，包括：
    利用第一高阶累积量判断第k视频帧中的所述目标是否被遮挡，所述第一高阶累积量对应于所述第k视频帧中的目标图像在子空间的第一重构误差；
    利用运动模型及所述第k视频帧中的目标状态信息预测并抽取得到多个预测粒子，所述第k视频帧中被遮挡与未被遮挡的所述目标的所述运动模型不同；
    分别计算每个所述预测粒子在第k+1视频帧中对应的预测图像块在所述子空间的重构误差；
    分别利用所述重构误差计算对应的每个所述预测粒子的重要性权值；
    利用第二高阶累积量判断所述第k视频帧中被遮挡的所述目标在所述第k+1视频帧中是否仍被遮挡，所述第二高阶累积量对应于似然图像块在所述子空间的第二重构误差，所述似然图像块是所述重要性权值最大的所述预测粒子对应的预测图像块；
    若所述第k视频帧和所述第k+1视频帧中所述目标均被遮挡，则将轨迹预测值作为所述第k+1视频帧中的目标图像，否则将所述似然图像块作为所述第k+1视频帧中的所述目标图像；
    利用所述第k+1视频帧中的目标图像获取所述第k+1视频帧中的目标状态信息。
  2. 根据权利要求1所述的方法，其特征在于，
    所述利用第一高阶累积量判断第k视频帧中的所述目标是否被遮挡包括：
    判断所述第一高阶累积量是否大于预设阈值；
    若所述第一高阶累积量大于预设阈值，则判定所述第k视频帧中的所述目标被遮挡，否则判定所述第k视频帧中的所述目标未被遮挡。
  3. 根据权利要求2所述的方法，其特征在于，
    所述第一重构误差e k为：
    Figure PCTCN2018092205-appb-100001
    其中，Z k为所述第k视频帧中的目标图像，U为所述子空间，
    Figure PCTCN2018092205-appb-100002
    为样本均值，f k(j)为所述第一重构误差中第j个像素点的灰度值，M k为所述第一重构误差中像素点的个数；
    所述第一高阶累积量C k为：
    Figure PCTCN2018092205-appb-100003
  4. 根据权利要求1所述的方法，其特征在于，
    所述利用运动模型及所述第k视频帧中的目标状态信息预测并抽取得到多个预测粒子包括：
    若所述第k视频帧中所述目标被遮挡，则利用随机游走模型及所述第k视频帧中的目标状态信息预测并抽取多个所述预测粒子，否则利用二阶自回归模型及所述第k视频帧中的目标状态信息预测并抽取多个所述预测粒子。
  5. 根据权利要求4所述的方法，其特征在于，
    所述第k视频帧中的目标状态信息X k＝{x k,y k,s k,α k}，其中x k和y k分别为所述第k视频帧中的目标图像中心点位置的x坐标和y坐标，s k和α k分别为所述第k视频帧中的目标图像与标准大小的比例以及宽高比；
    所述随机游走模型的状态转移概率p(X k+1|X k)为：
    p(X k+1|X k)＝N(X k+1|X k,Ψ)   (3)
    其中，X k+1为所述第k+1视频帧中的目标状态信息，N()为正态分布，ψ为对角化协方差矩阵，ψ的对角元素分别为
    Figure PCTCN2018092205-appb-100004
    δ x表示中心点位置x坐标的标准差，δ y表示中心点位置y坐标的标准差，δ s表示与标准大小的比例的标准差，δ α表示与标准大小的宽高比的标准差。
    所述二阶自回归模型为：
    X k+1-X k＝X k-X k-1+W k+1  (4)
    其中，W k+1为高斯白噪声。
  6. 根据权利要求1所述的方法，其特征在于，
    第i个所述预测粒子
    Figure PCTCN2018092205-appb-100005
    在第k+1视频帧中对应的预测图像块
    Figure PCTCN2018092205-appb-100006
    在所述子空间的重构误差
    Figure PCTCN2018092205-appb-100007
    为：
    Figure PCTCN2018092205-appb-100008
    其中，U为所述子空间，
    Figure PCTCN2018092205-appb-100009
    为样本均值，
    Figure PCTCN2018092205-appb-100010
    为所述重构误差
    Figure PCTCN2018092205-appb-100011
    中第j个像素点的灰度值，
    Figure PCTCN2018092205-appb-100012
    为所述重构误差
    Figure PCTCN2018092205-appb-100013
    中像素点的个数；
    第i个所述预测粒子的重要性权值
    Figure PCTCN2018092205-appb-100014
    满足
    Figure PCTCN2018092205-appb-100015
    Figure PCTCN2018092205-appb-100016
    为第i个所述预测粒子
    Figure PCTCN2018092205-appb-100017
    的观测似然函数，具体为：
    Figure PCTCN2018092205-appb-100018
    其中，δ表示标准正态分布的标准差，||·||为二范数。
  7. 根据权利要求1所述的方法，其特征在于，
    所述利用第二高阶累积量判断所述第k视频帧中被遮挡的所述目标在所述第k+1视频帧中是否仍被遮挡包括：
    在所述目标在所述第k视频帧中被遮挡的情况下，利用所述第二重构误差计算所述第二高阶累积量；
    判断所述第二高阶累积量是否大于预设阈值；
    若所述第二高阶累积量大于预设阈值，则判定所述第k+1视频帧中的所述目标被遮挡，否则判定所述第k+1视频帧中的所述目标未被遮挡。
  8. 根据权利要求7所述的方法，其特征在于，
    所述第二重构误差e k+1,max为：
    Figure PCTCN2018092205-appb-100019
    其中，Z k+1,max为所述似然图像块，U为所述子空间，
    Figure PCTCN2018092205-appb-100020
    为样本均值，f k+1,max(j)为所述第二重构误差中第j个像素点的灰度值，M k+1,max为所述第二重构误差中像素点的个数；
    所述第二高阶累积量C k+1,max为：
    Figure PCTCN2018092205-appb-100021
  9. 根据权利要求1所述的方法，其特征在于，进一步包括：
    所述目标在所述第k+1视频帧中未被遮挡的情况下，至少利用所述第k+1视频帧中的目标图像增量更新所述子空间和样本均值。
  10. 根据权利要求9所述的方法，其特征在于，
    更新前的所述子空间U对应的训练图像集为A＝{Z 1,Z 2,…,Z n}，n为所述训练图像集A中图像的数量，所述训练图像集A的均值为所述样本均值
    Figure PCTCN2018092205-appb-100022
    所述训练图像集A的中心数据矩阵为
    Figure PCTCN2018092205-appb-100023
    所述中心数据矩阵
    Figure PCTCN2018092205-appb-100024
    奇异值分解得到所述子空间U和特征向量Σ，新增图像集为B＝{Z n+1,Z n+2,…,Z n+m}，m为所述新增图像集B中图像的数量，所述新增图像集B包括所述k+1视频帧中的目标图像，m和n为大于或等于1的整数；
    更新后的所述子空间U'为：
    Figure PCTCN2018092205-appb-100025
    其中
    Figure PCTCN2018092205-appb-100026
    是对联合矩阵R进行奇异值分解得到的：
    Figure PCTCN2018092205-appb-100027
    所述联合矩阵R为：
    Figure PCTCN2018092205-appb-100028
    Figure PCTCN2018092205-appb-100029
    其中Orth()为执行正交化；
    Figure PCTCN2018092205-appb-100030
    其中
    Figure PCTCN2018092205-appb-100031
    为所述新增图像集B的均值；
    更新后的所述样本均值
    Figure PCTCN2018092205-appb-100032
    为：
    Figure PCTCN2018092205-appb-100033
    其中f为遗忘因子。
  11. 一种基于高阶累积量的目标跟踪装置，其特征在于，包括至少一个处理器，单独或协同工作，所述处理器用于执行指令以实现如权利要求1-10中任一项所述的方法。
  12. 一种可读存储介质，存储有指令，其特征在于，所述指令被执行时实现如权利要求1-10中任一项所述的方法。
PCT/CN2018/092205 2018-06-21 2018-06-21 基于高阶累积量的目标跟踪方法、装置及存储介质 WO2019241963A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/092205 WO2019241963A1 (zh) 2018-06-21 2018-06-21 基于高阶累积量的目标跟踪方法、装置及存储介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/092205 WO2019241963A1 (zh) 2018-06-21 2018-06-21 基于高阶累积量的目标跟踪方法、装置及存储介质

Publications (1)

Publication Number Publication Date
WO2019241963A1 true WO2019241963A1 (zh) 2019-12-26

Family

ID=68983138

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/092205 WO2019241963A1 (zh) 2018-06-21 2018-06-21 基于高阶累积量的目标跟踪方法、装置及存储介质

Country Status (1)

Country Link
WO (1) WO2019241963A1 (zh)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106296733A (zh) * 2016-08-04 2017-01-04 中国科学院长春光学精密机械与物理研究所 一种图像中目标跟踪方法
CN106355204A (zh) * 2016-08-31 2017-01-25 武汉大学 基于贝叶斯学习和增量子空间学习的目标跟踪方法

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106296733A (zh) * 2016-08-04 2017-01-04 中国科学院长春光学精密机械与物理研究所 一种图像中目标跟踪方法
CN106355204A (zh) * 2016-08-31 2017-01-25 武汉大学 基于贝叶斯学习和增量子空间学习的目标跟踪方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
LI, CHUNLAN: "Research on Particle Filtering Pedestrian Target Tracking Algorithm Based on Occlusion Detection", TECHNOLOGY, CHINA MASTER'S THESES FULL-TEXT DATABASE, 31 July 2017 (2017-07-31), ISSN: 1674-0246 *

Similar Documents

Publication Publication Date Title
WO2020107716A1 (zh) 目标图像分割方法、装置及设备
US9495600B2 (en) People detection apparatus and method and people counting apparatus and method
US9098740B2 (en) Apparatus, method, and medium detecting object pose
Ozyildiz et al. Adaptive texture and color segmentation for tracking moving objects
Ma et al. Towards a universal model for cross-dataset crowd counting
Cheng et al. A novel dynamic system in the space of SPD matrices with applications to appearance tracking
US9805285B2 (en) State estimation apparatus, state estimation method, and integrated circuit
Bernier et al. Fast nonparametric belief propagation for real-time stereo articulated body tracking
CN106780567B (zh) 一种融合颜色和梯度直方图的免疫粒子滤波扩展目标跟踪方法
CN111178261A (zh) 一种基于视频编码技术的人脸检测加速方法
Guan et al. Multi-pose human head detection and tracking boosted by efficient human head validation using ellipse detection
CN109166138B (zh) 基于高阶累积量的目标跟踪方法、装置及存储介质
Wang et al. Hamiltonian Monte Carlo estimator for abrupt motion tracking
Dutta et al. Weighted singular value thresholding and its application to background estimation
Ng et al. New models for real-time tracking using particle filtering
WO2019241963A1 (zh) 基于高阶累积量的目标跟踪方法、装置及存储介质
CN107767401B (zh) 基于核相关滤波的红外目标实时跟踪方法及装置
Li et al. Moving target tracking via particle filter based on color and contour features
Sahoo et al. Automated human tracking using advanced mean shift algorithm
Tian et al. Multi-scale human pose tracking in 2D monocular images
Chen et al. A generalized mean shift tracking algorithm
Lu et al. Person tracking with partial occlusion handling
Chu et al. Research of kernel particle filtering target tracking algorithm based on multi-feature fusion
Wang et al. An improved camshift-based particle filter algorithm for face tracking
Dhassi et al. IMM-PF visual tracking based on multi-feature fusion

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18922999

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 08.04.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 18922999

Country of ref document: EP

Kind code of ref document: A1