WO2015161697A1 - 应用于人机交互的运动物体跟踪方法及系统 - Google Patents
应用于人机交互的运动物体跟踪方法及系统 Download PDFInfo
- Publication number
- WO2015161697A1 WO2015161697A1 PCT/CN2015/071828 CN2015071828W WO2015161697A1 WO 2015161697 A1 WO2015161697 A1 WO 2015161697A1 CN 2015071828 W CN2015071828 W CN 2015071828W WO 2015161697 A1 WO2015161697 A1 WO 2015161697A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- tracking
- frame image
- threshold
- optical flow
- target
- Prior art date
Links
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
Definitions
- the present application relates to video images, and in particular, to a moving object tracking method and system for human-computer interaction.
- gesture-based device human-computer interaction has become very common. Based on the touch screen mode, people are using various gestures such as clicking, sliding, twisting, zooming, etc. every day. , iPad, Windows Phone or Android-based terminal devices for manipulation and entertainment.
- Such a large-scale user habit change is difficult to predict 5 to 10 years before the physical keyboard is used to realize human-computer interaction.
- the most important thing is that it is more intuitive and convenient, from home entertainment to enterprise applications. It has a long-lasting impact on all aspects of people's lifestyles in the future, and is an inevitable trend in the future technological innovation and upgrading of consumer electronics products.
- Kinect core module integrates the module into a set-top box or smart TV far beyond the affordability of ordinary consumers.
- the application provides a moving object tracking method and system applied to human-computer interaction.
- the present application provides a moving object tracking method applied to human-computer interaction, including:
- performing the LK optical flow tracking on the target area in the current frame image to obtain a tracking result frame specifically includes:
- a point pair that is successfully tracked is selected according to the first threshold.
- the first threshold is determined according to a relationship between an initial tracking point of the image of the previous frame and a tracking point obtained after two LK optical flow tracking, and the successful tracking is selected according to the first threshold.
- Point to point including:
- a pair of points whose error is less than the first threshold is selected.
- the acquiring an image and determining an initial position of the target in the current frame image includes:
- the initial position of the target is determined by skin color detection, motion block detection, and classifier detection.
- the calculating the credibility of the detection result, determining the final position of the target and the initial information for the next frame image LK optical flow tracking according to the credibility specifically includes:
- the detection result is taken as the final position of the target, and is used as initial information for the next frame image LK optical flow tracking;
- the tracking result is taken as the final position of the target, and is used as initial information of the next frame image LK optical flow tracking.
- the present application provides a moving object tracking system applied to human-computer interaction, including a fine detection module, a tracking module, a coarse detection module, and an analysis module;
- the fine detecting module is configured to collect an image and determine an initial position of the target in the current frame image
- the tracking module is configured to perform LK optical flow tracking on a target area in the current frame image to obtain a tracking result frame;
- the coarse detection module is configured to detect the current frame image by using a training classifier to obtain a detection result frame;
- the analyzing module is configured to calculate a credibility of the detection result, and determine, according to the credibility, a final position of the target and initial information used for the next frame image LK optical flow tracking.
- the tracking module includes a tracking unit and a selection unit
- the tracking unit is configured to track from the previous frame image to the current frame image by using an LK optical stream, and then backtrack from the tracked current frame image to the previous frame image by using an LK optical stream;
- the selecting unit is configured to determine a first threshold according to a relationship between an initial tracking point of the previous frame image and a tracking point obtained after two LK optical flow tracking, and select a successful tracking according to the first threshold Point to point.
- the selecting unit is further configured to calculate an error between an initial tracking point of the image of the previous frame and a tracking point obtained after two LK optical flow tracking, and the average value of the error is taken as the first
- the threshold is selected as a pair of points whose error is less than the first threshold.
- the fine detection module determines the initial position of the target specifically by skin color detection, motion block detection, and classifier detection.
- the analysis module includes a calculation unit and an analysis unit
- the calculating unit is configured to obtain the credibility by dividing an overlapping area of the tracking result frame and the detection result frame by an area of the tracking result frame;
- the analyzing unit is configured to determine a second threshold, when the credibility is greater than or equal to the second threshold, when the credibility is greater than or equal to the second threshold, the detection result is used as the target Final position, and as initial information for the next frame image LK optical flow tracking; when the reliability is less than the second threshold, the tracking result is taken as the final position of the target, and The initial information of the LK optical flow tracking of one frame of image.
- the LK optical flow tracking is performed on the current frame image to obtain a tracking result frame, and the current frame image is detected by the training classifier, and a detection result box is obtained, and the reliability of the detection result is calculated.
- the reliability determines the initial information of the next frame of image.
- the present application can use the target detection method in different situations in different situations, and combine detection and tracking to realize real-time and efficient tracking of dynamic targets. This application is a general method and is applicable to various human-computer interaction scenarios. Real-time tracking of moving objects without the need for wearable devices reduces hardware costs.
- the position of the gesture is determined by fine detection and the gesture information is judged, and then the image position information obtained by the fine detection is used as an initial state to track it, and finally the target tracking result of the moving object is combined.
- the result of the coarse detection is judged by the credibility, thereby correcting the position of the gesture target for the gesture target tracking of the next frame image.
- the application combines the methods of fine detection, coarse detection and LK optical flow tracking, which not only reduces the requirements on the system hardware, but also satisfies the real-time and accurate tracking.
- the application is applied to products such as set-top boxes or smart televisions, and the application is further reduced. cost.
- FIG. 1 is a flowchart of an embodiment of a moving object tracking method applied to human-computer interaction in an embodiment
- FIG. 2 is a flow chart of another embodiment of a moving object tracking method applied to human-computer interaction in the present application
- FIG. 3 is a schematic diagram of a coarse detection and a fine detection process of a moving object tracking method applied to human-computer interaction of the present application;
- FIG. 4 is a flow chart of tracking detection in a moving object tracking method applied to human-computer interaction of the present application
- FIG. 5 is a flowchart of a final result of tracking obtained in a moving object tracking method applied to human-computer interaction of the present application
- FIG. 6 is a schematic structural diagram of a moving object tracking system applied to human-computer interaction in an embodiment of the present application
- FIG. 7 is a schematic structural diagram of a moving object tracking system applied to human-computer interaction in an embodiment of the present application.
- Embodiment 1 is a diagrammatic representation of Embodiment 1:
- an embodiment of the method for tracking a moving object applied to human-computer interaction of the present application includes the following steps:
- Step 102 Acquire an image to determine an initial position of the target in the current frame image.
- Step 104 Perform LK optical flow tracking on the target area in the current frame image to obtain a tracking result frame.
- Step 106 Perform detection on the current frame image by using a training classifier to obtain a detection result frame.
- Step 108 Calculate the credibility of the detection result box, and determine the final position of the target and the initial information for the next frame image LK optical flow tracking according to the credibility.
- step 102 the initial position of the current frame image is determined, which is implemented by a fine detection method.
- Fine detection includes various detection methods such as skin color detection, motion block detection, and classifier detection.
- the moving object area and related information are determined by fine detection, and the detected image is used as an initial state for the next detection.
- the detected object can be a moving object such as a gesture or the like. Use fine detection to get the ROI of the gesture (Region Of Interest, region of interest).
- the detection is performed using a line classifier, which is called coarse detection.
- the image is detected by using the classifier obtained by Haar feature + AdaBoost training. Determine the specific location of the gesture and gesture information.
- step 104 uses the method of calculating the LK optical flow twice to obtain the information of the LK optical flow tracking point, and calculates the tracking error point by calculating the forward and backward error information, and estimates with the remaining tracking points.
- the location of the target in the current frame Specifically include:
- Step 1042 Tracking from the previous frame image to the current frame image by the LK optical stream.
- Step 1044 Backtracking from the tracked current frame image to the previous frame image through the LK optical stream.
- Step 1046 Determine a first threshold according to a relationship between an initial tracking point of the previous frame image and a tracking point obtained after two LK optical flow tracking.
- the error between the initial tracking point of the previous frame image and the tracking point obtained after two LK optical flow tracking is calculated, and the average value of all the errors is calculated, and the average value of the error is taken as the first threshold.
- the first threshold can also be set as needed, such as setting an empirical value or calculating by other methods.
- Step 1048 Select a point pair that is successfully tracked according to the first threshold. Specifically, a pair of points whose error is smaller than the first threshold may be selected.
- the position information of the target in the image of the previous frame is used as an initial state in which some points are uniformly selected as needed.
- 100 points can be selected, and 10*10 image blocks are taken as the center of each point, and the LK optical flow method is used to calculate the point A corresponding to the current frame image.
- Position B secondly, using the LK optical flow method again, the position B of the point in the current frame tracked in the previous direction is backtracked to obtain the position C in the previous frame; then, the calculations are similar to A and C respectively.
- Such a point-to-point error is taken as a pair of points that are less than or equal to the mean of the error as a correct pair of points; finally, the correct position is tracked to calculate the position and information transformation matrix, thereby obtaining the tracking position.
- step 108 specifically includes:
- Step 1082 The area of the overlap between the tracking result box and the detection result frame is divided by the area of the tracking result frame to obtain credibility
- Step 1084 Determine a second threshold; the second threshold may be determined according to requirements.
- the second threshold may take an empirical value, such as 0.65.
- Step 1086 When the reliability is greater than or equal to the second threshold, the detection result is used as the final location of the target, and is used as initial information for the next frame image LK optical flow tracking;
- Step 1088 When the reliability is less than the second threshold, the tracking result is taken as the final position of the target, and is used as the initial information of the next frame image LK optical flow tracking.
- FIG. 5 is a flow chart showing the final tracking results in a particular embodiment.
- Embodiment 2 is a diagrammatic representation of Embodiment 1:
- an embodiment of a moving object tracking system applied to human-computer interaction of the present application includes a fine detection module, a tracking module, a coarse detection module, and an analysis module.
- the fine detection module is configured to acquire an image to determine an initial position of the target in the current frame image;
- the tracking module is configured to perform LK optical flow tracking on the target area in the current frame image to obtain a tracking result frame;
- the coarse detection module is configured to pass the current frame image
- the training classifier performs detection to obtain a detection result frame;
- the analysis module is configured to calculate the credibility of the detection result, and determine a final position of the target according to the credibility and an optical flow tracking for the next frame image LK Initial information.
- the tracking module includes a tracking unit and a selection unit.
- the tracking unit is configured to track the image from the previous frame to the current frame image through the LK optical stream, and then backtrack from the tracked current frame image to the previous frame image through the LK optical stream; the selecting unit is used to initialize the image according to the previous frame.
- the selecting unit is further configured to calculate an error between an initial tracking point of the previous frame image and a tracking point obtained after two LK optical flow tracking, and use an average value of the error as the first threshold. A pair of points whose error is less than the first threshold.
- the fine detection module determines the initial position of the target specifically by skin tone detection, motion block detection, and classifier detection.
- the analysis module includes a calculation unit and an analysis unit; the calculation unit is configured to obtain the reliability by using the overlap area of the tracking result frame and the detection result frame by the area of the tracking result frame; the analysis unit is configured to determine the a second threshold, when the reliability is greater than or equal to the second threshold, the detection result is used as a final position of the target, and is used as initial information for the next frame image LK optical flow tracking; At the second threshold, the tracking result is taken as the final position of the target, and is used as the initial information of the next frame image LK optical flow tracking.
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
本申请公开了一种应用于人机交互的运动物体跟踪方法，包括采集图像，确定当前帧图像中目标的初始位置；对所述当前帧图像中的目标区域进行LK光流跟踪得到跟踪结果框；对所述当前帧图像通过训练分类器进行检测，得到检测结果框；计算所述检测结果的可信度，根据所述可信度确定所述目标的最终位置及用于下一帧图像LK光流跟踪的初始信息。本申请还公开了一种应用于人机交互的运动物体跟踪系统。本申请可在不同情况下，有区分地使用目标检测方法，并将检测与跟踪相结合来实现动态目标的实时高效跟踪，本申请为一种通用方法，适用于各种人机交互的场景，无需使用穿戴式设备，即可完成运动物体的实时跟踪，又降低了硬件成本。
Description
本申请涉及视频图像，尤其涉及一种应用于人机交互的运动物体跟踪方法及系统。
近年来随着智能手机和平板电脑的普及，基于手势的设备人机交互已经非常普遍，基于触摸屏的模式，人们每天都在通过点击、滑动、扭转、缩放等多种手势动作对诸如iPhone
、iPad、Windows
Phone或基于Android系统的终端设备进行操控、娱乐。这样大规模用户习惯的改变，在习惯物理键盘实现人机交互的5至10年前是难以预计的，最重要的是它以更直观、更便利的方式，从家庭娱乐到企业应用的方方面面，对未来人们的生活方式的各方面产生了长久深远的影响，是未来消费电子产品技术革新与升级的必然趋势。
2012年美国NMC Horizon
报告显示，人机交互在未来的4到5年，将会在现有基于触摸屏模式的交互方式，将以智能电视、家庭娱乐、手机应用等消费电子品为媒介，向基于图像、语音和MEMS传感器的多种模式偕同发展，彻底改变未来人们的生活、工作、娱乐中的信息交互方式。
目前，国内外的研发状态主要呈现以下特点：
1）多数同类产品针对特定行业应用或本公司配套设备而设计，而非通用模块。例如Google手套是与Google
Glass搭配的产品，任天堂的游戏手柄及微软Kinect都是配套其公司游戏主机的辅助装置，尽管也可以在其它领域得以应用，但仍有很大的制约与局限。
2）多数类似系统价格昂贵，与现有通用电子消费平整和成本过高，在价格上难以大众普及。例如Kinect核心模块，把该模块集成到机顶盒或智能电视之中远超普通消费者的承受能力。
本申请提供一种应用于人机交互的运动物体跟踪方法及系统。
根据本申请的第一方面，本申请提供一种应用于人机交互的运动物体跟踪方法，包括：
采集图像，确定当前帧图像中目标的初始位置；
对所述当前帧图像中的目标区域进行LK光流跟踪得到跟踪结果框；
对所述当前帧图像通过训练分类器进行检测，得到检测结果框；
计算所述检测结果的可信度，根据所述可信度确定所述目标的最终位置及用于下一帧图像LK光流跟踪的初始信息。
上述方法中，所述对所述当前帧图像中的目标区域进行LK光流跟踪得到跟踪结果框，具体包括：
通过LK光流从所述上一帧图像跟踪到所述当前帧图像；
再从跟踪到的所述当前帧图像通过LK光流反跟踪到所述上一帧图像；
根据所述上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的关系，确定第一阈值；
根据所述第一阈值选取成功跟踪的点对。
上述方法中，所述根据所述上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的关系，确定第一阈值，根据所述第一阈值选取成功跟踪的点对，具体包括：
计算所述上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的误差，将所述误差的平均值作为第一阈值；
选取所述误差小于所述第一阈值的点对。
上述方法中，所述采集图像，确定当前帧图像中目标的初始位置，具体包括：
通过肤色检测、运动块检测和分类器检测确定所述目标的初始位置。
上述方法中，所述计算所述检测结果的可信度，根据所述可信度确定所述目标的最终位置及用于下一帧图像LK光流跟踪的初始信息，具体包括：
使用所述跟踪结果框与所述检测结果框的重叠面积除以所述跟踪结果框的面积，得到所述可信度；
确定第二阈值；
当所述可信度大于等于所述第二阈值时，将所述检测结果作为所述目标的最终位置，并作为用于下一帧图像LK光流跟踪的初始信息；
当所述可信度小于所述第二阈值时，将所述跟踪结果作为所述目标的最终位置，并作为下一帧图像LK光流跟踪的初始信息。
根据本申请的第二方面，本申请提供一种应用于人机交互的运动物体跟踪系统，包括细检测模块、跟踪模块、粗检测模块和分析模块；
所述细检测模块用于采集图像，确定当前帧图像中目标的初始位置；
所述跟踪模块用于对所述当前帧图像中的目标区域进行LK光流跟踪得到跟踪结果框；
所述粗检测模块用于对所述当前帧图像通过训练分类器进行检测，得到检测结果框；
所述分析模块用于计算所述检测结果的可信度，根据所述可信度确定所述目标的最终位置及用于下一帧图像LK光流跟踪的初始信息。
上述系统中，所述跟踪模块包括跟踪单元和选择单元；
所述跟踪单元用于通过LK光流从所述上一帧图像跟踪到所述当前帧图像，再从跟踪到的所述当前帧图像通过LK光流反跟踪到所述上一帧图像；
所述选择单元用于根据所述上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的关系，确定第一阈值，并根据所述第一阈值选取成功跟踪的点对。
上述系统中，所述选择单元还用于计算所述上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的误差，将所述误差的平均值作为第一阈值，选取所述误差小于所述第一阈值的点对。
上述系统中，所述细检测模块具体通过肤色检测、运动块检测和分类器检测确定所述目标的初始位置。
上述系统中，所述分析模块包括计算单元和分析单元；
所述计算单元用于使用所述跟踪结果框与所述检测结果框的重叠面积除以所述跟踪结果框的面积，得到所述可信度；
所述分析单元用于确定第二阈值，在所述可信度大于等于所述第二阈值时，当所述可信度大于等于所述第二阈值时，将所述检测结果作为所述目标的最终位置，并作为用于下一帧图像LK光流跟踪的初始信息；在所述可信度小于所述第二阈值时，将所述跟踪结果作为所述目标的最终位置，并作为下一帧图像LK光流跟踪的初始信息。
由于采用了以上技术方案，使本申请具备的有益效果在于：
(1)
在本申请的具体实施方式中，包括对当前帧图像进行LK光流跟踪得到跟踪结果框，对当前帧图像通过训练分类器进行检测，得到检测结果框，计算检测结果的可信度，根据可信度确定下一帧图像的初始信息。本申请可在不同情况下，有区分地使用目标检测方法，并将检测与跟踪相结合来实现动态目标的实时高效跟踪，本申请为一种通用方法，适用于各种人机交互的场景，无需使用穿戴式设备，即可完成运动物体的实时跟踪，又降低了硬件成本。
(2)在本申请的具体实施方式中，利用细检测确定手势的位置及判断手势信息，再以细检测得到的图像位置信息为初始状态对它进行跟踪，最后对运动物体目标跟踪结果，结合粗检测结果，进行可信度判断,从而修正手势目标所在位置,用于下一帧图像的手势目标跟踪。本申请通过细检测、粗检测、LK光流跟踪相结合的方法，既降低了对系统硬件的要求，又能满足实时准确的跟踪，将本申请用于机顶盒或智能电视等产品，进一步降低了成本。
图1为本申请的应用于人机交互的运动物体跟踪方法在一种实施方式中的流程图；
图2为本申请的应用于人机交互的运动物体跟踪方法在另一种实施方式中的流程图；
图3是本申请的应用于人机交互的运动物体跟踪方法粗检测和细检测流程示意图；
图4是本申请的应用于人机交互的运动物体跟踪方法中跟踪检测的流程图；
图5是本申请的应用于人机交互的运动物体跟踪方法中得到跟踪的最终结果的流程图；
图6为本申请的应用于人机交互的运动物体跟踪系统在一种实施方式中的结构示意图；
图7为本申请的应用于人机交互的运动物体跟踪系统在一种实施方式中的结构示意图。
下面通过具体实施方式结合附图对本申请作进一步详细说明。
实施例一：
如图1、图2所示，本申请的应用于人机交互的运动物体跟踪方法，其一种实施方式，包括以下步骤：
步骤102：采集图像，确定当前帧图像中目标的初始位置。
步骤104：对当前帧图像中的目标区域进行LK光流跟踪得到跟踪结果框。
步骤106：对当前帧图像通过训练分类器进行检测，得到检测结果框。
步骤108：计算检测结果框的可信度，根据可信度确定目标的最终位置及用于下一帧图像LK光流跟踪的初始信息。
如图3所示，步骤102中，确定当前帧图像的初始位置，通过细检测的方法实现。细检测包括肤色检测、运动块检测和分类器检测等多种检测方法。通过细检测确定运动物体区域及相关信息，以检测到的图像作为下一步检测的初始状态。检测的对象可以是运动物体，如手势等。利用细检测来获取手势所在ROI
（Region Of Interest ，感兴趣区域)。
本步骤106中，使用线下分类器进行检测，称为粗检测。利用Haar特征+AdaBoost训练得到的分类器，对图像进行手势检测。确定手势的具体位置及手势信息。
在一种实施方式中，步骤104利用计算两次LK光流的方法来获得LK光流跟踪点的信息，通过计算前后向误差信息来过滤误差较大的跟踪点，用余下的跟踪点来估计当前帧中目标所在位置。具体包括：
步骤1042：通过LK光流从上一帧图像跟踪到当前帧图像。
步骤1044：再从跟踪到的当前帧图像通过LK光流反跟踪到上一帧图像。
步骤1046：根据上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的关系，确定第一阈值。
计算上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的误差，计算所有误差的平均值，将误差的平均值作为第一阈值。第一阈值也可以根据需要来设置，如设置一个经验值，或是通过其他方法计算。
步骤1048：根据第一阈值选取成功跟踪的点对。具体可选取误差小于所述第一阈值的点对。
首先，利用目标在前一帧图像中的位置信息作为初始状态，在该位置区域中根据需要均匀选取一些点。如图4所示，在本实施方式中，可选取100个点，并以每个点为中心，取10*10的图像块，利用LK光流法，来计算该点A对应于当前帧图像的位置B；其次，再次使用LK光流法，对前面跟踪到的当前帧中该点的位置B进行反向跟踪，得到其在前一帧中的位置C；然后，分别计算类似A、C这样的100个点对之间误差，取小于等于误差均值的点对作为跟踪正确的点对；最后，利用这些跟踪正确的点来计算位置及信息变换矩阵，从而得到跟踪位置。
本申请的应用于人机交互的运动物体跟踪方法，步骤108具体包括：
步骤1082：使用跟踪结果框与检测结果框的重叠面积除以跟踪结果框的面积，得到可信度；
步骤1084：确定第二阈值；第二阈值可以根据需要来确定，在本实施方式中，第二阈值可以取一个经验值，如0.65。
步骤1086：当可信度大于等于第二阈值时，将检测结果作为所述目标的最终位置，并作为用于下一帧图像LK光流跟踪的初始信息；
步骤1088：当可信度小于第二阈值时，将跟踪结果作为所述目标的最终位置，并作为下一帧图像LK光流跟踪的初始信息。
如果可信度小于第二阈值，则取跟踪结果为最终跟踪结果；否则，取粗检测结果为最终跟踪结果。如果没有跟踪结果，则使用细检测，取检测结果为最终跟踪结果。图5为在一种具体实施方式中得到最终的跟踪结果的流程图。
实施例二：
如图6所示，本申请的应用于人机交互的运动物体跟踪系统，其一种实施方式，包括细检测模块、跟踪模块、粗检测模块和分析模块。细检测模块用于采集图像，确定当前帧图像中目标的初始位置；跟踪模块用于对当前帧图像中的目标区域进行LK光流跟踪得到跟踪结果框；粗检测模块用于对当前帧图像通过训练分类器进行检测，得到检测结果框；分析模块用于计算所述检测结果的可信度，根据所述可信度确定所述目标的最终位置及用于下一帧图像LK光流跟踪的初始信息。
如图7所示，跟踪模块包括跟踪单元和选择单元。跟踪单元用于通过LK光流从上一帧图像跟踪到当前帧图像，再从跟踪到的当前帧图像通过LK光流反跟踪到上一帧图像；选择单元用于根据上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的关系，确定第一阈值，并根据第一阈值选取成功跟踪的点对。
在一种实施方式中，选择单元还用于计算上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的误差，将误差的平均值作为第一阈值，选取误差小于第一阈值的点对。
在一种实施方式中，细检测模块具体通过肤色检测、运动块检测和分类器检测确定目标的初始位置。
在一种实施方式中，分析模块包括计算单元和分析单元；计算单元用于使用跟踪结果框与检测结果框的重叠面积除以跟踪结果框的面积，得到可信度；分析单元用于确定第二阈值，在可信度大于等于所述第二阈值时，将所述检测结果作为所述目标的最终位置，并作为用于下一帧图像LK光流跟踪的初始信息；在可信度小于第二阈值时，将跟踪结果作为所述目标的最终位置，并作为下一帧图像LK光流跟踪的初始信息。
以上内容是结合具体的实施方式对本申请所作的进一步详细说明，不能认定本申请的具体实施只局限于这些说明。对于本申请所属技术领域的普通技术人员来说，在不脱离本申请构思的前提下，还可以做出若干简单推演或替换。
Claims (10)
- 一种应用于人机交互的运动物体跟踪方法，其特征在于，包括：采集图像，确定当前帧图像中目标的初始位置；对所述当前帧图像中的目标区域进行LK光流跟踪得到跟踪结果框；对所述当前帧图像通过训练分类器进行检测，得到检测结果框；计算所述检测结果的可信度，根据所述可信度确定所述目标的最终位置及用于下一帧图像LK光流跟踪的初始信息。
- 如权利要求1所述的应用于人机交互的运动物体跟踪方法，其特征在于，所述对所述当前帧图像中的目标区域进行LK光流跟踪得到跟踪结果框，具体包括：通过LK光流从所述上一帧图像跟踪到所述当前帧图像；再从跟踪到的所述当前帧图像通过LK光流反跟踪到所述上一帧图像；根据所述上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的关系，确定第一阈值；根据所述第一阈值选取成功跟踪的点对。
- 如权利要求2所述的应用于人机交互的运动物体跟踪方法，其特征在于，所述根据所述上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的关系，确定第一阈值，根据所述第一阈值选取成功跟踪的点对，具体包括：计算所述上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的误差，将所述误差的平均值作为第一阈值；选取所述误差小于所述第一阈值的点对。
- 如权利要求１所述的应用于人机交互的运动物体跟踪方法，其特征在于，所述采集图像，确定当前帧图像中目标的初始位置，具体包括：通过肤色检测、运动块检测和分类器检测确定所述目标的初始位置。
- 如权利要求１所述的应用于人机交互的运动物体跟踪方法，其特征在于，所述计算所述检测结果的可信度，根据所述可信度确定所述目标的最终位置及用于下一帧图像LK光流跟踪的初始信息，具体包括：使用所述跟踪结果框与所述检测结果框的重叠面积除以所述跟踪结果框的面积，得到所述可信度；确定第二阈值；当所述可信度大于等于所述第二阈值时，将所述检测结果作为所述目标的最终位置，并作为用于下一帧图像LK光流跟踪的初始信息；当所述可信度小于所述第二阈值时，将所述跟踪结果作为所述目标的最终位置，并作为下一帧图像LK光流跟踪的初始信息。
- 一种应用于人机交互的运动物体跟踪系统，其特征在于，包括细检测模块、跟踪模块、粗检测模块和分析模块；所述细检测模块用于采集图像，确定当前帧图像中目标的初始位置；所述跟踪模块用于对所述当前帧图像中的目标区域进行LK光流跟踪得到跟踪结果框；所述粗检测模块用于对所述当前帧图像通过训练分类器进行检测，得到检测结果框；所述分析模块用于计算所述检测结果的可信度，根据所述可信度确定所述目标的最终位置及用于下一帧图像LK光流跟踪的初始信息。
- 如权利要求6所述的应用于人机交互的运动物体跟踪系统，其特征在于，所述跟踪模块包括跟踪单元和选择单元；所述跟踪单元用于通过LK光流从所述上一帧图像跟踪到所述当前帧图像，再从跟踪到的所述当前帧图像通过LK光流反跟踪到所述上一帧图像；所述选择单元用于根据所述上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的关系，确定第一阈值，并根据所述第一阈值选取成功跟踪的点对。
- 如权利要求7所述的应用于人机交互的运动物体跟踪系统，其特征在于，所述选择单元还用于计算所述上一帧图像的初始跟踪点与经过两次LK光流跟踪后得到的跟踪点之间的误差，将所述误差的平均值作为第一阈值，选取所述误差小于所述第一阈值的点对。
- 如权利要求6所述的应用于人机交互的运动物体跟踪系统，其特征在于，所述细检测模块具体通过肤色检测、运动块检测和分类器检测确定所述目标的初始位置。
- 如权利要求6所述的应用于人机交互的运动物体跟踪系统，其特征在于，所述分析模块包括计算单元和分析单元；所述计算单元用于使用所述跟踪结果框与所述检测结果框的重叠面积除以所述跟踪结果框的面积，得到所述可信度；所述分析单元用于确定第二阈值，在所述可信度大于等于所述第二阈值时，当所述可信度大于等于所述第二阈值时，将所述检测结果作为所述目标的最终位置，并作为用于下一帧图像LK光流跟踪的初始信息；在所述可信度小于所述第二阈值时，将所述跟踪结果作为所述目标的最终位置，并作为下一帧图像LK光流跟踪的初始信息。
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201410172131.8 | 2014-04-25 | ||
| CN201410172131.8A CN103985137B (zh) | 2014-04-25 | 2014-04-25 | 应用于人机交互的运动物体跟踪方法及系统 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2015161697A1 true WO2015161697A1 (zh) | 2015-10-29 |
Family
ID=51277095
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2015/071828 WO2015161697A1 (zh) | 2014-04-25 | 2015-01-29 | 应用于人机交互的运动物体跟踪方法及系统 |
Country Status (2)
| Country | Link |
|---|---|
| CN (1) | CN103985137B (zh) |
| WO (1) | WO2015161697A1 (zh) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109028227A (zh) * | 2018-09-29 | 2018-12-18 | 佛山市云米电器科技有限公司 | 智能油烟机及其移动人体检测方法 |
| CN109726665A (zh) * | 2018-12-25 | 2019-05-07 | 中国科学院合肥物质科学研究院 | 一种基于动态轨迹分析的农业害虫检测方法 |
| CN110738668A (zh) * | 2019-09-29 | 2020-01-31 | 南京佑驾科技有限公司 | 一种智能控制远光灯的方法、系统及交通工具 |
| CN111462180A (zh) * | 2020-03-30 | 2020-07-28 | 西安电子科技大学 | 基于与或图aog的目标跟踪方法 |
| CN112037259A (zh) * | 2020-08-27 | 2020-12-04 | 北京极智嘉科技有限公司 | 动态目标的跟踪系统及方法 |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103985137B (zh) * | 2014-04-25 | 2017-04-05 | 深港产学研基地 | 应用于人机交互的运动物体跟踪方法及系统 |
| CN104835182A (zh) * | 2015-06-03 | 2015-08-12 | 上海建炜信息技术有限公司 | 摄像机上实现动态目标实时跟踪的方法 |
| CN106780563A (zh) * | 2017-01-22 | 2017-05-31 | 王恒升 | 一种带回测光流的图像特征点跟踪法 |
| CN107403444B (zh) * | 2017-09-12 | 2023-07-18 | 深圳市中联讯科技有限公司 | 识别系统 |
| CN108334811B (zh) * | 2017-12-26 | 2021-06-04 | 大唐软件技术股份有限公司 | 一种人脸图像处理方法及装置 |
| CN110363790B (zh) * | 2018-04-11 | 2024-06-14 | 北京京东尚科信息技术有限公司 | 目标追踪方法、装置和计算机可读存储介质 |
| CN109345564A (zh) * | 2018-07-30 | 2019-02-15 | 深圳市艾为智能有限公司 | 一种解决由于自相似特征产生光流场与运动场不符的方法 |
| CN111627046A (zh) * | 2020-05-15 | 2020-09-04 | 北京百度网讯科技有限公司 | 目标部位跟踪方法、装置、电子设备和可读存储介质 |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102270348A (zh) * | 2011-08-23 | 2011-12-07 | 中国科学院自动化研究所 | 基于视频流的对变形手势进行跟踪的方法 |
| CN103426008A (zh) * | 2013-08-29 | 2013-12-04 | 北京大学深圳研究生院 | 基于在线机器学习的视觉人手跟踪方法及系统 |
| CN103593679A (zh) * | 2012-08-16 | 2014-02-19 | 北京大学深圳研究生院 | 一种基于在线机器学习的视觉人手跟踪方法 |
| CN103985137A (zh) * | 2014-04-25 | 2014-08-13 | 北京大学深圳研究院 | 应用于人机交互的运动物体跟踪方法及系统 |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103034851B (zh) * | 2012-12-24 | 2015-08-26 | 清华大学深圳研究生院 | 自学习的基于肤色模型的手部跟踪装置及方法 |
-
2014
- 2014-04-25 CN CN201410172131.8A patent/CN103985137B/zh active Active
-
2015
- 2015-01-29 WO PCT/CN2015/071828 patent/WO2015161697A1/zh active Application Filing
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102270348A (zh) * | 2011-08-23 | 2011-12-07 | 中国科学院自动化研究所 | 基于视频流的对变形手势进行跟踪的方法 |
| CN103593679A (zh) * | 2012-08-16 | 2014-02-19 | 北京大学深圳研究生院 | 一种基于在线机器学习的视觉人手跟踪方法 |
| CN103426008A (zh) * | 2013-08-29 | 2013-12-04 | 北京大学深圳研究生院 | 基于在线机器学习的视觉人手跟踪方法及系统 |
| CN103985137A (zh) * | 2014-04-25 | 2014-08-13 | 北京大学深圳研究院 | 应用于人机交互的运动物体跟踪方法及系统 |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109028227A (zh) * | 2018-09-29 | 2018-12-18 | 佛山市云米电器科技有限公司 | 智能油烟机及其移动人体检测方法 |
| CN109726665A (zh) * | 2018-12-25 | 2019-05-07 | 中国科学院合肥物质科学研究院 | 一种基于动态轨迹分析的农业害虫检测方法 |
| CN109726665B (zh) * | 2018-12-25 | 2022-10-25 | 中国科学院合肥物质科学研究院 | 一种基于动态轨迹分析的农业害虫检测方法 |
| CN110738668A (zh) * | 2019-09-29 | 2020-01-31 | 南京佑驾科技有限公司 | 一种智能控制远光灯的方法、系统及交通工具 |
| CN111462180A (zh) * | 2020-03-30 | 2020-07-28 | 西安电子科技大学 | 基于与或图aog的目标跟踪方法 |
| CN111462180B (zh) * | 2020-03-30 | 2023-03-24 | 西安电子科技大学 | 基于与或图aog的目标跟踪方法 |
| CN112037259A (zh) * | 2020-08-27 | 2020-12-04 | 北京极智嘉科技有限公司 | 动态目标的跟踪系统及方法 |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103985137A (zh) | 2014-08-13 |
| CN103985137B (zh) | 2017-04-05 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2015161697A1 (zh) | 应用于人机交互的运动物体跟踪方法及系统 | |
| WO2016176959A1 (zh) | 一种基于眼球追踪技术的显示屏多屏控制方法及系统 | |
| WO2011013976A2 (en) | Method and apparatus for controlling electronic device using user interaction | |
| WO2014030902A1 (en) | Input method and apparatus of portable device | |
| WO2014135023A1 (zh) | 一种智能终端的人机交互方法及系统 | |
| WO2013105765A1 (en) | Display apparatus and controlling method thereof | |
| WO2018076912A1 (zh) | 一种虚拟场景调整方法及头戴式智能设备 | |
| WO2020253453A1 (zh) | 图像切换方法、装置、电子设备及存储介质 | |
| EP2649511A2 (en) | 3 dimensional (3d) display system of responding to user motion and user interface for the 3d display system | |
| WO2017088486A1 (zh) | 一种移动终端提醒方法及系统 | |
| WO2015137742A1 (en) | Display apparatus and controlling method thereof | |
| WO2016086433A1 (zh) | 一种触控基板、终端及提高触摸精度的方法 | |
| WO2017211056A1 (zh) | 一种移动终端的单手操作方法、及系统 | |
| WO2019045244A1 (ko) | 시각 대화를 통해 객체의 위치를 알아내기 위한 주의 기억 방법 및 시스템 | |
| WO2014178693A1 (ko) | 복수의 디바이스를 매칭시키는 방법, 그리고 상기 매칭이 가능하게끔 하여 주는 디바이스 및 서버 시스템 | |
| WO2011053059A2 (en) | Electronic apparatus for proximity sensing | |
| WO2017065450A1 (en) | Display apparatus and controlling method thereof | |
| WO2015196878A1 (zh) | 一种电视虚拟触控方法及系统 | |
| WO2011081354A2 (en) | Korean input method and apparatus using touch screen, and portable terminal including key input apparatus | |
| WO2016095362A1 (zh) | 显示终端及其屏幕的显示方法 | |
| WO2013066095A1 (ko) | 얼굴 검출 방법, 장치 및 이 방법을 실행하기 위한 컴퓨터 판독 가능한 기록 매체 | |
| WO2021244650A1 (zh) | 控制方法、装置、终端及存储介质 | |
| CN103914305A (zh) | 一种移动终端上自由控制应用的方法及系统 | |
| WO2012034469A1 (zh) | 基于手势的人机交互方法及系统、计算机存储介质 | |
| WO2014117655A1 (zh) | 触控屏幕的点击操作结构 |
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: 15782429 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 15782429 Country of ref document: EP Kind code of ref document: A1 |