WO2018028102A1 - 一种仿记忆引导的模式识别方法 - Google Patents

一种仿记忆引导的模式识别方法 Download PDF

Info

Publication number
WO2018028102A1
WO2018028102A1 PCT/CN2016/109001 CN2016109001W WO2018028102A1 WO 2018028102 A1 WO2018028102 A1 WO 2018028102A1 CN 2016109001 W CN2016109001 W CN 2016109001W WO 2018028102 A1 WO2018028102 A1 WO 2018028102A1
Authority
WO
WIPO (PCT)
Prior art keywords
motion
memory
pattern recognition
sequence
segment
Prior art date
Application number
PCT/CN2016/109001
Other languages
English (en)
French (fr)
Inventor
陈哲
王志坚
胡文才
王鑫
Original Assignee
河海大学
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 河海大学 filed Critical 河海大学
Priority to AU2016418565A priority Critical patent/AU2016418565B2/en
Priority to NZ750970A priority patent/NZ750970A/en
Priority to GB1902335.7A priority patent/GB2567595B/en
Priority to US16/323,113 priority patent/US10860891B2/en
Publication of WO2018028102A1 publication Critical patent/WO2018028102A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/20—Movements or behaviour, e.g. gesture recognition
    • G06V40/23—Recognition of whole body movements, e.g. for sport training
    • G06V40/25—Recognition of walking or running movements, e.g. gait recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/20—Analysis of motion
    • G06T7/215—Motion-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/62—Extraction of image or video features relating to a temporal dimension, e.g. time-based feature extraction; Pattern tracking
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/40—Scenes; Scene-specific elements in video content
    • G06V20/49—Segmenting video sequences, i.e. computational techniques such as parsing or cutting the sequence, low-level clustering or determining units such as shots or scenes
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/103—Static body considered as a whole, e.g. static pedestrian or occupant recognition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10016—Video; Image sequence
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/20—Special algorithmic details
    • G06T2207/20048—Transform domain processing
    • G06T2207/20056—Discrete and fast Fourier transform, [DFT, FFT]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/30—Subject of image; Context of image processing
    • G06T2207/30242—Counting objects in image

Definitions

  • the invention relates to a pattern recognition method, in particular to a pattern recognition and motion detection method in a video sequence guided by a memory.
  • the human visual perception system has a prominent advantage in pattern recognition. More advanced memory and selection systems enable human vision to be quickly focused on motion and continuous motion changes. This mechanism is implemented by the guidance of global timing features in short-term memory and long-term memory, called memory guidance. Based on this mechanism, pattern recognition for continuous motion states is achieved by merging spatial positions and motion trajectories of targets at different times. For example, in Figure 1, the human perception system describes the scene as “a girl running from the left to the right", “a lot of vehicles are moving in the opposite direction", and “two people come together.” It can be found that each description except “running girl”, “traveling car” and “pedestrian” emphasizes mode information such as a motion trajectory in a continuous time series up to the current time.
  • a temporal filter that simulates the center of the visual periphery, a motion saliency model based on spectral differences, etc., but these models cannot describe the motion trajectory in the entire time series, and the ability to suppress motion noise is weak, and the result is weak. There are more noise interferences.
  • the basic basis of the invention is that the changes produced by the motion are not only related to the samples of the adjacent fields but also to the global context.
  • long-term memory segments need to be introduced into the pattern recognition process, and the state of change of motion throughout the time series can be obtained in one detection process.
  • the present invention provides a time-sequence pattern recognition method in a memory-guided manner, which can obtain a target motion state in an entire time series in a single recognition process, and It can solve the problem that noise is difficult to suppress in motion detection of complex natural scenes.
  • a pattern recognition method based on imitation memory guidance comprising the following steps:
  • Step 1 The imitation memory calling mechanism and its process divide the historical time series and combine it with the current time frame to become a segment sequence as a primitive for pattern recognition. For a video sequence we can get multiple sequence sequences, independent of each other, without overlap, and processed in parallel.
  • Step 2 Simulate the motion saliency mechanism and its process, extract the motion saliency in each segment sequence, and detect the motion information in the short-term sequence.
  • the visual motion saliency model is used to detect the motion changes occurring in the sequence of the segment.
  • Step 3 imitating the memory decay mechanism and its process, weighting the motion information, and weighting the motion information in all the sequence sequences, outputting the current moment motion information and the motion trajectory in the entire time series, and synthesizing it as a pattern recognition result.
  • the weighted cumulative fusion method can improve the accuracy of the current time motion detection, and can obtain the motion trajectory in the entire memory time sequence interval, and comprehensively obtain the overall pattern recognition result.
  • the present invention takes the fragment sequence as the basis Pattern recognition is performed by performing motion detection and fusion processing.
  • This strategy can accurately detect the motion information of the current time and generate the trajectory information of the motion state change in the entire time series, and can calibrate the time information of the trajectory and motion, and has the ability to suppress the motion noise.
  • Figure 1 is a schematic diagram of pattern recognition of memory guidance
  • FIG. 2 is a flow chart of pattern recognition according to an embodiment of the present invention, (a) inputting a frame image of a current time in a video sequence, (b) motion detection in short-term memory, and (c) motion detection in long-term memory, ( d) motion detection in the frame image at the current time, (e) motion state in the entire time series;
  • FIG. 3 is a schematic diagram of time series segmentation in an embodiment of the present invention.
  • the implementation is as shown in Fig. 2, a pattern recognition method based on memory-guided.
  • the present invention mainly relies on four biological discoveries: 1 short-term memory can guide the recognition of recent movement mutations; 2 long-term memory can guide the recognition of stable movement changes; 3 memory-based motion saliency mechanisms are concentrated In the short-term memory, long-term memory, the part of the movement changes, ignoring the irrelevant background and sporadic noise information; 4 through the interaction of memory and motion saliency mechanism can describe the current time information and the entire time series in memory Motion state and state change (motion trajectory).
  • pattern recognition in short-term memory should be able to calibrate motion information for those mutations that have recently occurred and suppress static background information (see Figure 2(b)).
  • pattern recognition in long-term memory should be able to calibrate motion information that is stable over a longer time range (see Figure 2(c)) and suppress sporadic motion noise.
  • the highest intensity pattern recognition results obtained are concentrated in those parts of the short-term/long-term memory that have undergone motion changes (Fig. 2(d)).
  • the pattern recognition method should be able to detect the trajectory of motion changes throughout the time series in memory to accurately describe the motion state and state changes (see Figure 2(e)).
  • the time series in one memory segment (memory storage is l) can be combined by x n and the sample in memory And get it. Therefore, for the point x n in the current moment, the corresponding sequence fragment can be constructed as:
  • x n is the point considered at the current time
  • x n-1 , x n-2 ,..., x nl ⁇ k is the sample in memory
  • k is the length of the time series (the minimum value is set to 16)
  • l is the amount of memory stored (the minimum value is set to 10).
  • motion information in each segment sequence is detected using a time-based Fourier transform-based visual motion saliency model.
  • the visual motion saliency model based on time Fourier transform considers that the fluctuation of the phase spectrum in the time series spectrum corresponds to the change of the sequence information in the time domain. Motion information can thus be detected by calculation of the phase spectrum. It mainly includes the following steps:
  • Step 1 Construct a sequence of fragments consisting of the current time point and the historical time sample:
  • Step two calculating the Fourier transform and the corresponding phase spectrum for the sequence of the segment:
  • step three the inverse Fourier transform is calculated for the obtained phase spectrum:
  • F and F -1 represent the Fourier transform and the inverse Fourier transform, respectively.
  • Step 4 performing threshold filtering on the inverse transform result of the phase spectrum, if If the value is greater than the threshold, the motion information appears at the corresponding position, otherwise there is no motion change:
  • T is the threshold and the typical value is set to:
  • the motion detection results in the plurality of sequence segments are fused to form a memory-guided motion detection result, which not only takes into account the short-term motion mutation but also considers the stable motion change over a long period of time.
  • This mechanism can be modeled as the accumulation of motion detection results in all 1 sequence segments:
  • E n is the accumulation of motion detection results.
  • the result obtained is shown in Fig. 4: in the obtained motion detection result map (Fig. 4), the point at which the intensity is the largest corresponds to the motion information at the current time, and the intensity is weakened with the continuation of time. Therefore, the motion detection result map E n can be divided by the method based on the gray histogram. By this calculation, it is possible to obtain the motion information S n for only the current time as the pattern recognition result of the current time.
  • the intensity of motion information in the human brain should diminish over time, known as memory attenuation.
  • the present invention employs a weighted approach to simulate this attenuation mechanism. For any sequence of segments, the intensity of the detected motion information is inversely proportional to the delay of the segment from the current time. This weight and attenuation can be calculated as:
  • the motion information in the current moment can be combined with the motion information of the historical moment to form a motion trajectory, and the present invention formalizes the simulation of the mechanism as the accumulation of the motion information after the attenuation:

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Multimedia (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Human Computer Interaction (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Psychiatry (AREA)
  • Social Psychology (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Medical Informatics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)
  • Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)

Abstract

一种仿记忆引导的模式识别方法，通过引入当前时刻之前记忆中的时间序列信息，以序列中的样本为参考进行模式识别，检测运动信息。模拟人脑中的记忆调用机制，以不同的记忆片段为基元，检测出多个记忆片段中的运动变化及与之相对应的运动状态；模拟记忆的衰减机制，组合多个片段中的运动检测结果，增强当前时刻中的运动变化信息并构成连续时间序列中的运动状态，形成当前时刻的运动检测结果，作为模式识别结果。本方法能够稳定、可靠地用于复杂条件下的模式识别及运动检测，具有较好的噪声抑制效果且运算效率较高。

Description

一种仿记忆引导的模式识别方法 技术领域
本发明涉及一种模式识别方法，具体的是指一种仿记忆引导的视频序列中模式识别及运动检测方法。
背景技术
人类的视觉感知***在模式识别上具有较为突出的优势。较为高级的记忆和选择***能够使人类的视觉快速的集中于运动和连续的运动变化上。这种机制是通过短时记忆和长时记忆中的全局时序特征引导而实现的，称之为记忆引导。基于该种机制，对于连续运动状态的模式识别是通过融合不同时刻目标的空间位置及运动轨迹而实现的。例如在图1中，人类感知***对场景的描述为：“一个女孩从左边向右边跑”，“许多车辆相向行驶”，以及“两个人走到了一起”。可以发现，除了“跑步的女孩”，“行驶的汽车”和“行人”之外每种描述都强调了在连续的时间序列中至当前时刻为止的运动轨迹等模式信息。然而，目前的机器视觉、图像处理及模式识别方法多是针对当前时刻目标静态的空间位置进行预测和检测、或是仅针对目标的运动轨迹进行拟合，尚无方法能够模拟人类记忆机制以片段序列为基元同时检测出当前目标的空间位置并能够回溯其运动轨迹等模式，如图1所示。
此外，对于真实、复杂场景中的模式识别还存在噪声难以抑制的问题。目前的机器视觉方法多是采用对序列中所有数据一次建模的方法以识别当前帧中运动模式的变化。在这一过程中，真实场景中大量的运动噪声被引入到结果中，并且这种运动噪声不仅数量大且难以抑制，严重影响到模式识别及运动检测结果的准确性。对于这一问题的解决，很多工作给我们有益的启发。例如，模拟视觉中央.周边差的时间滤波器、基于频谱差分的运动显著性模型等，但是这些模型均无法实现对整个时间序列中运动轨迹的描述，且对运动噪声的抑制能力较弱，结果中存在较多的噪声干扰。
受益于最新生物学研究的发展，发现人类视觉的模式识别在很大程度上要依赖于记忆中的历史经验，短时或长时记忆中的全局上下文信息使得模式识别更加便捷、高效。这一发现强调了全局长下文信息对模式识别的重要作用。这种重要 作用主要体现在它不仅能够准确的检测运动变化模式并得到运动轨迹，还能够抑制由于相机抖动等所导致的运动噪声。因此，需要对这种机制进行建模，发明一种全新的模式识别模型以准确的检测当前时刻的运动并标定运动轨迹，在最终的模式识别结果中，同时准确检测出当前时刻和记忆中历史时刻的运动变化、运动轨迹。该发明的基本依据在于运动所产生的变化不仅同时域邻接的样本有关还同全局的上下文有关。因而，除了短时记忆片段，长时记忆片段也需要引入到模式识别过程中，能够在一次检测过程中得到运动在整个时间序列中的变化状态。
发明内容
发明目的：针对现有技术中存在的问题，本发明提供一种仿记忆引导的时间序列中模式识别方法，该模式识别方法能够在一次识别过程中得到在整个时间序列中的目标运动状态，并能够解决复杂自然场景运动检测中噪声难以抑制的问题。
技术方案：一种仿记忆引导的模式识别方法，包括如下步骤：
步骤一：仿记忆调用机制及其过程，分割历史时间序列并同当前时刻帧组合成为片段序列作为模式识别的基元。对于一个视频序列我们可以获得多个片段序列，彼此间相互独立，无重叠，且并行处理。
步骤二：仿视觉运动显著性机制及其过程，提取每个片段序列中的运动显著性，检测得到该短时序列中的运动信息。对于每一个记忆片段中的时间序列，采用视觉运动显著性模型检测该片段序列中所出现的运动变化。
步骤三：仿记忆衰退机制及其过程，对运动信息进行加权，并加权融合所有片段序列中的运动信息，输出当前时刻运动信息及整个时间序列中的运动轨迹，综合作为模式识别结果。
对于每个片段中的运动检测结果，考虑到记忆片段其同当前时刻的时延，认为时延越大的片段序列中运动检测结果同当前时刻的时间相关性越弱，所赋相应权重值越小；反之，认为时延越小的片段序列中运动检测结果同当前时刻的时间相关性越强，所赋相应权重值越大。通过加权累积的融合的方式能够提高对当前时刻运动检测的准确性，并能够得到整个记忆时序列区间中的运动轨迹，综合得到整体的模式识别结果。
相比较一般模式识别方法对时间序列的整体建模，本发明以片段序列为基元 进行运动检测并融合的处理方式进行模式识别。这种策略能够准确检测出当前时刻的运动信息并生成整个时间序列中运动状态变化的轨迹信息，且能够标定轨迹、运动的时间信息，并兼有对运动噪声的抑制能力。
附图说明
图1是记忆引导的模式识别示意图；
图2是本发明实施例的模式识别的流程图，(a)输入视频序列中当前时刻的帧图像，(b)短时记忆中的运动检测，(c)长时记忆中的运动检测，(d)当前时刻中帧图像中的运动检测，(e)整个时间序列中的运动状态；
图3是本发明实施例中时间序列分割的示意图；
图4是本发明实施例中积累的运动检测及模式识别结果。
具体实施方式
下面结合具体实施例，进一步阐明本发明，应理解这些实施例仅用于说明本发明而不用于限制本发明的范围，在阅读了本发明之后，本领域技术人员对本发明的各种等价形式的修改均落于本申请所附权利要求所限定的范围。
实施例如图2所示，一种仿记忆引导的模式识别方法。本发明的提出主要依赖于四个生物学发现：①短时记忆能够引导识别近期所出现的运动突变；②长时记忆能够引导识别稳定出现的运动变化；③基于记忆的运动显著性机制多集中于短时记忆、长时记忆中均发生运动变化的部分，忽视无关的背景及零星的噪声信息；④通过记忆和运动显著性机制的交互能够描述当前时刻中的运动信息及记忆中整个时间序列中的运动状态及状态变化(运动轨迹)。根据发现①，在短时记忆中的模式识别应能够标定出那些近期所出现突变的运动信息，并抑制静态的背景信息(如图2(b))。根据发现②，在长时记忆中的模式识别应能够标定出较长时间范围内所稳定出现的运动信息(如图2(c))并抑制零星的运动噪声。根据发现③，所得到的强度最高的模式识别结果集中于那些短时\长时记忆中均发生运动变化的部分(如图2(d))。根据发现④，模式识别方法应能够检测到在记忆中整个时间序列中运动变化的轨迹，以准确的描述运动状态及状态变化(如图2(e))。
首先，根据发现①、②若一个点被识别为发生了运动模式的变化，那么该点应该在短时记忆和长时记忆片段中均检测到了稳定的运动变化。为了模拟这种机制首先对记忆中的时间序列进行分割，并将同一空间位置、时间上连续的样本 排列组成片段序列。对于当前时刻的点，仅在它同所有记忆中的样本存在差异时才认为其位置处发生了运动。分割过程如图3所示。对于当前时刻中的点xn，若该点的信息同记忆中存储的样本信息存在差异那么认为该点位置处发生运动变化。具体而言，若xt为t时刻所考察的点，当前时刻t＝n，l个记忆片段(记忆的存储量为l)中的时间序列可通过组合xn及记忆中的样本
Figure PCTCN2016109001-appb-000001
而得到。因此，对于当前时刻中的点xn，相应的序列片段可以构造为：
Figure PCTCN2016109001-appb-000002
其中，
Figure PCTCN2016109001-appb-000003
为片段序列，xn为当前时刻所考察的点，xn-1，xn-2，…，xn-l×k为记忆中的样本，k为时间序列的长度(最小值设置为16)，l为记忆的存储量(最小值设置为10)。
至此，完成了对记忆中所有时间序列的分割。
在每个序列片段中，采用基于时间傅里叶变换的视觉运动显著性模型检测每个片段序列中的运动信息。基于时间傅里叶变换的视觉运动显著性模型认为时间序列频谱中相位谱的波动对应于时域中序列信息的变化。因而可以通过对相位谱的计算检测运动信息。主要包括以下步骤：
步骤一，构造当前时刻点和历史时刻样本所组成的片段序列：
Figure PCTCN2016109001-appb-000004
步骤二，对于该片段序列计算其傅里叶变换及相应的相位谱：
Figure PCTCN2016109001-appb-000005
步骤三，对所得的相位谱计算其反傅里叶变换：
Figure PCTCN2016109001-appb-000006
其中，F和F-1分别表示傅里叶变换及反傅里叶变换，
Figure PCTCN2016109001-appb-000007
代表序列片段
Figure PCTCN2016109001-appb-000008
的相位谱，
Figure PCTCN2016109001-appb-000009
为相位谱的反傅里叶变换结果，g(t)为一维高斯滤波器(典型值方 差σ＝5)。为了准确地检测运动信息的同时抑制背景中的运动噪声，需要进一步对
Figure PCTCN2016109001-appb-000010
进行阈值滤波。
步骤四，对相位谱的反变换结果进行阈值滤波，如果
Figure PCTCN2016109001-appb-000011
的值大于阈值则相应位置处出现运动信息，否则认为无运动变化：
Figure PCTCN2016109001-appb-000012
其中，T为阈值，典型值设置为：
Figure PCTCN2016109001-appb-000013
其中，
Figure PCTCN2016109001-appb-000014
和
Figure PCTCN2016109001-appb-000015
分别为
Figure PCTCN2016109001-appb-000016
的均值和方差。
随后，根据发现③，将多个序列片段中的运动检测结果融合，形成记忆引导的运动检测结果，不仅考虑到了短时间内的运动突变还考虑了长时间内稳定的运动变化。这种机制可以形式化建模为所有l个序列片段中的运动检测结果的累积：
Figure PCTCN2016109001-appb-000017
其中，En为运动检测结果的累积。所得到的结果如图4所示：在所得到的运动检测结果图中(图4)强度最大的点对应于当前时刻的运动信息，强度随着时间的延续而减弱。因此，可以通过基于灰度直方图的方法对运动检测结果图En进行分割。通过这种计算，可以得到仅针对当前时刻的运动信息Sn，作为当前时刻的模式识别结果。
至此，完成了对当前时刻运动信息的检测，能够识别得到当前时刻的模式。
根据发现④，运动信息在人脑中的强度应随着时间的推移而减弱，称之为记忆衰减。本发明采用加权的方法来模拟这种衰减机制。对于任意片段序列，所检测到运动信息对应的强度反比于该片段距离当前时刻的时延。这种权重及衰减可以计算为：
Figure PCTCN2016109001-appb-000018
其中，
Figure PCTCN2016109001-appb-000019
为第i个片段中运动检测结果所对应的权重，α为调制参数，取值范 围为0＜α＜1。由于
Figure PCTCN2016109001-appb-000020
的强度随着时延的增大而降低，使其能够标定运动发生的时间。
此外，根据发现④，当前时刻中的运动信息可以同历史时刻的运动信息组合形成运动轨迹，本发明形式化模拟这种机制为衰减后运动信息的累积：
Figure PCTCN2016109001-appb-000021
其中，
Figure PCTCN2016109001-appb-000022
为运动轨迹图。为了抑制轨迹生成过程及运动信息衰减过程中所引入的噪声。将原始运动检测累积结果同衰减后的运动检测累积结果相乘：
Figure PCTCN2016109001-appb-000023
其中，
Figure PCTCN2016109001-appb-000024
为噪声抑制后的轨迹，所得到的结果如图2(e)所示。
至此，完成了记忆中所有时刻运动轨迹的拟合和表征，完成了对历史时刻运动轨迹及运动状态等模式的识别。

Claims (3)

  1. 一种仿记忆引导的模式识别方法，其特征在于，包括如下步骤：
    步骤一：仿记忆调用机制及其过程，分割历史时间序列并同当前时刻帧组合成为片段序列作为模式识别的基元；
    步骤二：仿视觉运动显著性机制及其过程，提取每个片段序列中的运动显著性，检测得到该短时序列中的运动信息；
    步骤三：仿记忆衰退机制及其过程，对运动信息进行加权，并加权融合所有片段序列中的运动信息，输出当前时刻运动信息及整个时间序列中的运动轨迹，综合作为模式识别结果。
  2. 如权利要求1所述的仿记忆引导的模式识别方法，其特征在于，仿记忆调用机制中基于片段序列的运动检测策略，既对于所存储的历史记忆进行分段，以记忆片段作为运动检测的基元；具体，对于视频序列中当前时刻t的图像帧，将1，2，…，t时刻的帧按照长度k进行分割，分别对分割出每个片段序列模拟视觉运行显著性机制检测片段序列中所出现的运动变换，作为该片段中的模式识别结果。
  3. 如权利要求1所述的仿记忆引导的模式识别方法，其特征在于，仿记忆衰退机制中，对于每个片段中的运动检测结果，考虑到记忆片段同当前时刻的时延，认为时延越大的片段序列中运动检测结果同当前时刻的时间相关性越弱，所赋相应权重值越小；反之，认为时延越小的片段序列中运动检测结果同当前时刻的时间相关性越强，所赋相应权重值越大；通过加权累积融合的方式能够提高对当前时刻运动检测的准确性，并能够得到整个记忆时序列区间中的运动轨迹，综合得到整体的模式识别结果。
PCT/CN2016/109001 2016-08-08 2016-12-08 一种仿记忆引导的模式识别方法 WO2018028102A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
AU2016418565A AU2016418565B2 (en) 2016-08-08 2016-12-08 Memory mimicry guided pattern recognition method
NZ750970A NZ750970A (en) 2016-08-08 2016-12-08 Memory-guide simulated pattern recognition method
GB1902335.7A GB2567595B (en) 2016-08-08 2016-12-08 Memory-guide simulated pattern recognition method
US16/323,113 US10860891B2 (en) 2016-08-08 2016-12-08 Memory-guide simulated pattern recognition method

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201610643658.3A CN106296736B (zh) 2016-08-08 2016-08-08 一种仿记忆引导的模式识别方法
CN2016106436583 2016-08-08

Publications (1)

Publication Number Publication Date
WO2018028102A1 true WO2018028102A1 (zh) 2018-02-15

Family

ID=57666794

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/109001 WO2018028102A1 (zh) 2016-08-08 2016-12-08 一种仿记忆引导的模式识别方法

Country Status (6)

Country Link
US (1) US10860891B2 (zh)
CN (1) CN106296736B (zh)
AU (1) AU2016418565B2 (zh)
GB (1) GB2567595B (zh)
NZ (1) NZ750970A (zh)
WO (1) WO2018028102A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114770511A (zh) * 2022-05-09 2022-07-22 上海傅利叶智能科技有限公司 基于物理触觉的机器人控制方法、装置及机器人

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107169406B (zh) * 2017-03-28 2020-06-30 中山大学 一种通过力台压力中心数据提取身体平衡振荡起始时间的方法
CN109256184A (zh) * 2018-07-30 2019-01-22 邓建晖 一种基于认知记忆识别与恢复方法及***
CN109271889B (zh) * 2018-08-29 2021-10-08 中山大学 一种基于双层lstm神经网络的动作识别方法
CN114897929B (zh) * 2022-05-31 2024-06-04 工业云制造(四川)创新中心有限公司 基于视觉降噪的机器人运动方法
CN116612125B (zh) * 2023-07-21 2023-09-29 山东庆葆堂生物科技有限公司 基于人工智能的食药品胶囊质量检测方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101311947A (zh) * 2008-06-12 2008-11-26 浙江大学 基于自然视频的实时智能监控方法
CN103793054A (zh) * 2014-01-17 2014-05-14 中南民族大学 一种模拟陈述性记忆过程的动作识别方法
CN104461000A (zh) * 2014-12-03 2015-03-25 北京航空航天大学 一种基于少量缺失信号的在线连续人体运动识别方法
US20160078287A1 (en) * 2014-08-29 2016-03-17 Konica Minola Laboratory U.S.A., Inc. Method and system of temporal segmentation for gesture analysis

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060018516A1 (en) * 2004-07-22 2006-01-26 Masoud Osama T Monitoring activity using video information
CN101271525B (zh) * 2008-04-10 2011-05-04 复旦大学 一种快速的图像序列特征显著图获取方法
US20110134245A1 (en) * 2009-12-07 2011-06-09 Irvine Sensors Corporation Compact intelligent surveillance system comprising intent recognition
US8649558B2 (en) * 2011-05-31 2014-02-11 Wisconsin Alumni Research Foundation Video processing with region-based warping
US8891009B2 (en) * 2011-08-29 2014-11-18 Futurewei Technologies, Inc. System and method for retargeting video sequences
US20130289824A1 (en) * 2012-04-30 2013-10-31 GM Global Technology Operations LLC Vehicle turn assist system and method
US10074015B1 (en) * 2015-04-13 2018-09-11 Google Llc Methods, systems, and media for generating a summarized video with video thumbnails
US9699371B1 (en) * 2016-03-29 2017-07-04 Sony Corporation Image processing system with saliency integration and method of operation thereof

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101311947A (zh) * 2008-06-12 2008-11-26 浙江大学 基于自然视频的实时智能监控方法
CN103793054A (zh) * 2014-01-17 2014-05-14 中南民族大学 一种模拟陈述性记忆过程的动作识别方法
US20160078287A1 (en) * 2014-08-29 2016-03-17 Konica Minola Laboratory U.S.A., Inc. Method and system of temporal segmentation for gesture analysis
CN104461000A (zh) * 2014-12-03 2015-03-25 北京航空航天大学 一种基于少量缺失信号的在线连续人体运动识别方法

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114770511A (zh) * 2022-05-09 2022-07-22 上海傅利叶智能科技有限公司 基于物理触觉的机器人控制方法、装置及机器人
CN114770511B (zh) * 2022-05-09 2023-06-23 上海傅利叶智能科技有限公司 基于物理触觉的机器人控制方法、装置及机器人

Also Published As

Publication number Publication date
US10860891B2 (en) 2020-12-08
GB201902335D0 (en) 2019-04-03
CN106296736B (zh) 2019-03-29
AU2016418565A1 (en) 2019-03-14
GB2567595A (en) 2019-04-17
GB2567595B (en) 2022-03-02
US20190164013A1 (en) 2019-05-30
CN106296736A (zh) 2017-01-04
NZ750970A (en) 2020-06-26
AU2016418565B2 (en) 2020-04-16

Similar Documents

Publication Publication Date Title
WO2018028102A1 (zh) 一种仿记忆引导的模式识别方法
Chung et al. An efficient hand gesture recognition system based on deep CNN
JP4208898B2 (ja) 対象物追跡装置および対象物追跡方法
CN106462976B (zh) 在异步传感器观测到的场景中追踪形状的方法
WO2017129020A1 (zh) 视频中人体行为识别的方法、装置和计算机存储介质
JP2009087090A (ja) 対象物追跡装置および対象物追跡方法
CN111027481B (zh) 基于人体关键点检测的行为分析方法及装置
KR20160087738A (ko) 이벤트 기반 센서를 이용한 객체의 검출 방법 및 장치
Manikandan et al. Video object extraction by using background subtraction techniques for sports applications
CN108491764A (zh) 一种视频人脸情绪识别方法、介质及设备
CN107945210A (zh) 基于深度学习和环境自适应的目标跟踪算法
CN110991278A (zh) 计算机视觉***的视频中人体动作识别方法和装置
CN112801019B (zh) 基于合成数据消除无监督车辆再识别偏差的方法及***
Liu et al. ACDnet: An action detection network for real-time edge computing based on flow-guided feature approximation and memory aggregation
CN109558790B (zh) 一种行人目标检测方法、装置及***
CN108830170A (zh) 一种基于分层特征表示的端到端目标跟踪方法
US8428369B2 (en) Information processing apparatus, information processing method, and program
CN110598646A (zh) 一种基于深度特征的无约束重复动作计数方法
CN108257148B (zh) 特定对象的目标建议窗口生成方法及其在目标跟踪的应用
Muthuswamy et al. Salient motion detection through state controllability
Zeng et al. Dynamic equilibrium module for action recognition
KR101777948B1 (ko) 엔트로피와 배경변화율 차이점을 이용한 주요 객체 자동 검출 장치
CN111105441B (zh) 一种受前面帧目标信息约束的相关滤波目标跟踪方法
Raghuveera et al. Comparative study of CAMSHIFT and RANSAC methods for face and eye tracking in real-time video
Sun Adaptive Multi-scale Fusion Siamese Network for Visual Tracking

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: 16912556

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 201902335

Country of ref document: GB

Kind code of ref document: A

Free format text: PCT FILING DATE = 20161208

ENP Entry into the national phase

Ref document number: 2016418565

Country of ref document: AU

Date of ref document: 20161208

Kind code of ref document: A

122 Ep: pct application non-entry in european phase

Ref document number: 16912556

Country of ref document: EP

Kind code of ref document: A1