CN107563342B - Pedestrian robust tracking method for field search and rescue of unmanned aerial vehicle - Google Patents
Pedestrian robust tracking method for field search and rescue of unmanned aerial vehicle Download PDFInfo
- Publication number
- CN107563342B CN107563342B CN201710835495.3A CN201710835495A CN107563342B CN 107563342 B CN107563342 B CN 107563342B CN 201710835495 A CN201710835495 A CN 201710835495A CN 107563342 B CN107563342 B CN 107563342B
- Authority
- CN
- China
- Prior art keywords
- particle
- particles
- state
- template
- pixels
- Prior art date
- Legal status (The legal status 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 status listed.)
- Expired - Fee Related
Links
Images
Landscapes
- Image Analysis (AREA)
Abstract
The invention provides a pedestrian robust tracking method for field search and rescue of unmanned aerial vehicles, which comprises the following steps: inputting a frame of image, and initializing a template and a particle set; propagating the sampled particles in the set of particles to satisfy a known probability distribution of importance; updating the weight of the particle by using a diversity measurement method; normalizing the weight of the particles and estimating a target state; the particles are screened and resampled. Experiments show that compared with the similar tracking method, the method has the advantages that the effective tracking is realized, and the tracking efficiency is greatly improved, so that the method has wide application value.
Description
Technical Field
The invention relates to the field of target tracking, in particular to a pedestrian robust tracking method for field search and rescue of unmanned aerial vehicles.
Background
At present, along with the diversification of national form of going out, the application demand of open-air search and rescue is constantly promoting, and the degree of difficulty of open-air search and rescue also constantly increases, and how in time effectual open-air search and rescue of developing becomes a problem that waits to solve urgently. In order to complete search and rescue tasks, a large number of professional search and rescue personnel are often required to be collected in a traditional field search mode, and omission is possibly caused while a large amount of manpower, material resources and time are consumed, so that the mode is expensive and low in efficiency. Therefore, it is very important to design an efficient and intelligent search and rescue system. The micro unmanned aerial vehicle has the characteristics of portability, flexibility and wide detection range, so that the unmanned aerial vehicle has a great number of advantages and a wide application prospect when being combined with the unmanned aerial vehicle to carry out field search and rescue.
A pedestrian robust tracking method for field search and rescue application of an unmanned aerial vehicle is a key technology for realizing field search and rescue of the unmanned aerial vehicle. However, due to the influences of environmental occlusion, view angle transformation, light change, rotation transformation, rigidity and non-rigid deformation of the observed target and the like which may occur in the tracking process, the unmanned aerial vehicle is easy to generate deviation in the process of distinguishing the observed target and lose the target, so that many mainstream target tracking methods cannot meet the application requirements. However, the idea of particle filtering can effectively meet the above challenges, so we will use the framework of particle filtering to construct the tracking method.
The traditional particle filter framework mainly comprises four stages, namely a particle state propagation stage, a particle weight updating stage, a particle weight normalization stage and a particle state estimation stage, wherein the four stages are mutually buckled and influenced layer by layer, and can estimate the expectation of the particle state under the condition of not providing state space hidden distribution, so that the traditional particle filter framework is widely applied to a nonlinear and non-Gaussian system. However, the particle filter framework performs simulation using a large number of particles, and thus has a problem of a large amount of calculation and high complexity. To solve this problem, we introduce a diversity metric approach in the construction of the tracking method.
The diversity measurement is an efficient and accurate template matching method. In this method, a good match has a high diversity and small amount of distortion, while a wrong match target has less diversity and large amount of distortion. In addition, the diversity measurement method also has the characteristics of nonparametric and unidirectional mapping, so that the method is suitable for improving the efficiency by using a region-of-interest extraction method, and the aim of improving the particle filtering performance is fulfilled.
However, a technique for performing target tracking using the characteristics of the diversity metric has not yet appeared in the art.
Disclosure of Invention
The invention aims to provide a particle filtering method based on diversity measurement, which can better complete the task of target tracking in the field search and rescue process of an unmanned aerial vehicle and can effectively track pedestrians moving in the field environment with higher precision.
The pedestrian robust tracking method for field search and rescue of unmanned aerial vehicles provided by the invention specifically comprises the following steps:
a pedestrian robust tracking method for field search and rescue of unmanned aerial vehicles is characterized by comprising the following steps:
step 1.1, read in initialization frame I0Acquiring a proper target from the target as a tracking template;
step 1.2, initializing and defining the state space of the target by utilizing the position and size information of the template to obtain the initial state X of the target0；
Step 1.3, according to the initial state X of the target0Randomly generating N particles to form an initialized particle set
Wherein the content of the first and second substances,
obeying a mean value of X0The variance is a Gaussian distribution of phi, and
then the weight value is normalized by the particles and is initialized to 1/N;
step 2.1, region-of-interest extraction, namely, each particle in the single view particle set is not considered any more, but the region formed by the particles is considered as a region-of-interest (ROI), and the region-of-interest is treated as a whole, specifically, the particle set at the time k is acquired
Fusing the particles to obtain a region of interest (ROI) at the K moment;
step 2.2, performing superpixel segmentation, namely extracting pixels of the region of interest and the template, and expressing the characteristics in a mode of replacing the pixels with the superpixels;
step 2.3, nearest neighbor matching; performing nearest neighbor matching on the super-pixels of the regions of interest and the super-pixels of the template, and finding the nearest neighbor of the super-pixels in the super-pixel set of the template for the super-pixels in each region of interest;
step 2.4, diversity measurement, namely utilizing the idea of nearest neighbor matching to carry out similarity measurement so as to realize the updating of the particle weight;
step 2.5, calculating the weight of the particles; uniformly substituting the similarity measurement results into a Gaussian distribution (taking the measured value as a mean value and the measured noise as a variance) for calculation, wherein the obtained result is the weight of the particle state
And 3, estimating the particle state, namely performing normalization operation on the weight of the particles, performing weighted summation by combining the states of the particles to obtain the state of the tracking target in the current frame, specifically the weight of the normalized particles, performing weighted summation by combining the particle states, wherein the summation result is the estimated target state:
step 4, resampling the particles, namely screening the particles from the angle of probability statistical distribution, so as to inhibit the degradation of the state weight of the particles, specifically: the weight of the accumulated particles
And selecting a random number m-R (0,1) in accordance with uniform distribution, and determining an order of arrival using the random number
Particles of established state
It is added into the particle set and initialized to have a state weight of 1/N.
In the above pedestrian robust tracking method for field search and rescue of unmanned aerial vehicle, the step 2 specifically implements the method including: defining an importance probability distribution satisfies the following constraints:
q(xk|x0:k-1,y1:k)＝q(xk|xk-1,yk)
wherein xkAnd ykRespectively representing the state value and the measured value of the particle at the time k, q (-) represents the probability distribution of importance, and
then represents the state noise of the system (which can be solved by the system state transition equation); it can be seen that the propagation process of the particle can be predicted according to the state transition equation, and the propagated particle can be obtained through the markov process; thus, in this embodiment, the particle state propagation will be done in two steps:
step 2.1, obtaining the particle set at the k-1 moment
And propagating the state of the particles to obtain a corresponding prediction result:
wherein v isk-1Obeying a Gaussian distribution with the mean value of 0 and the variance of phi as state noise;
step 2.2, the particles obtained by the transmission are combined to obtain a state set of the system at the time k
Wherein the state of the particles is represented by
The propagation is obtained and the weight of the particle is calculated from
The transfer is obtained.
In the above pedestrian robust tracking method for field search and rescue of unmanned aerial vehicle, the step 2.2 specifically implements the method comprising:
step A, dividing an image into grids, and determining the center of each grid as an initialization clustering center;
b, distributing the pixel points in the grid to the nearest cluster center by using a k-means method;
c, adjusting the clustering center according to the characteristics of the pixel points, so that each pixel can be distributed to the neighbor of the original clustering center;
and step D, iteratively performing the operations from the step A to the step C, so that all pixels can converge on a specific clustering center.
In the above pedestrian robust tracking method for field search and rescue of unmanned aerial vehicle, the step 2.3 specifically implements the method comprising: performing nearest neighbor matching on the super-pixels of the regions of interest and the super-pixels of the template, and finding the nearest neighbor of the super-pixels in the super-pixel set of the template for the super-pixels in each region of interest:
NN(q,PROI)＝argminp∈Pd(pA,qA)
wherein d (-) represents the pixel distance, PROIRepresenting a superpixel set resulting from the segmentation of the region of interest, Q representing a superpixel set resulting from the segmentation of the template region, and pA、qAThe characteristics of the superpixels in set P, Q are respectively represented.
In the pedestrian robust tracking method for field search and rescue of unmanned aerial vehicle, step 2.4 is implemented as a method packetComprises the following steps: defining a template superpixel point set as
And the target superpixel point set is
The diversity metric method is formulated as:
w(qj)＝NNW(qj,P)-k(NN(qj,P))
wherein the content of the first and second substances,
is a regularization factor, k (p)i) The fraction of the nearest neighbor corresponding to the target superpixel found in the template superpixel set is determined.
The invention has the following advantages: 1. aiming at the problems of large calculation amount and high complexity of the traditional particle filtering, a diversity measurement method is introduced, and a particle filtering method based on diversity measurement is provided; 2. the characteristics of diversity measurement nonparametric and one-way mapping are combined, a region-of-interest extraction method is introduced, and pixels which may be used by particles are obtained in advance by means of pixel extraction, so that repeated operation in the weight value updating process is avoided; 3. aiming at the problem of an unbalanced point set existing in diversity measurement, a superpixel segmentation method is introduced, so that particles can be expressed by a series of superpixels in a normalized mode, and the problem of the unbalanced point set is solved in a hidden mode;
drawings
Fig. 1 is an overall block diagram of a tracking method.
Fig. 2 is a flow of implementing the updating of the particle weight.
Fig. 3a is a graph of the results of comparison of effects (distance accuracy) with the same type of method.
Fig. 3b is a graph of the results of comparison of the effects with the same type of method (overlay accuracy).
Fig. 4a is a graph of the results of comparison of effects (distance accuracy) with the same type of method.
FIG. 4b is a graph showing the result of comparison of the effects of the same methods (overlay accuracy)
FIG. 5 is a graph showing the results of comparing the efficiency with the same method.
Detailed Description
The following further describes an implementation mode and a basic principle of a pedestrian robust tracking method facing field rescue of the unmanned aerial vehicle with reference to the attached drawings.
And step 1, initializing. The initialization of the video image mainly comprises the initialization of a template and the initialization of a particle set, the main purpose is to provide a traceable target and an initial sampling particle for a particle filter, and the implementation points are briefly as follows:
1) read-in initialization frame I0Acquiring a proper target from the target as a tracking template;
2) initializing and defining the state space of the target by utilizing the position and size information of the template to obtain the initial state X of the target0；
3) According to the initial state X of the target0Randomly generating N particles to form an initialized particle set
Wherein the content of the first and second substances,
obeying a mean value of X0The variance is a Gaussian distribution of phi, and
then the weight value is normalized by the particles and is initialized to 1/N;
and 2, propagating the particle state, namely propagating the sampling particles in the particle set to meet a known importance probability distribution. Here, we define that the importance probability distribution satisfies the following constraint:
q(xk|x0:k-1,y1:k)＝q(xk|xk-1,yk)
wherein xkAnd ykRespectively representing the state value and the measured value of the particle at the time k, q (-) represents the probability distribution of importance, and
it represents the state noise of the system (which can be solved by the system state transition equation). It can be seen that the propagation process of the particle can be predicted according to the state transition equation, and the propagated particle can be obtained through the markov process. Thus, in this embodiment, the particle state propagation will be done in two steps:
1) obtaining a set of particles at time k-1
And propagating the state of the particles to obtain a corresponding prediction result:
wherein v isk-1Obeying a Gaussian distribution with the mean value of 0 and the variance of phi as state noise;
2) combining the particles obtained by propagation to obtain a state set of the system at the k moment
Wherein the state of the particles is represented by
The propagation is obtained and the weight of the particle is calculated from
The transmission is obtained;
and 3, updating the weight of the particle, and updating the weight of the particle by using a diversity measurement method. The specific implementation flow is as follows:
1) and extracting a region of interest. Obtaining a set of particles at time k
And fusing the particles to obtain a region of interest (ROI) at the K moment. In combination, this provides several benefits:
● the region formed by the particles is regarded as the interested region, so that the repeated calculation in the similarity measurement process can be greatly reduced;
● the size of the super pixel point set can be reasonably planned by determining the region of interest of the particle set, and the reasonable segmentation of the particles can be realized;
● after the strategy of region of interest extraction is adopted, the tracking effect can be improved by increasing the number of particles without worrying about the reduction of the tracking efficiency;
2) and (4) super-pixel segmentation. Extracting the superpixels in the region of interest and the template by using an SLIC (narrow line Carrier) method, and combining to form a superpixel set of the region of interest and the template, wherein the implementation key points are briefly as follows:
● dividing the image into grids, and determining the center of the grids as an initialization clustering center;
● distributing the pixel points in the grid to the nearest cluster center by using a k-means method;
●, adjusting the clustering center according to the characteristics of the pixel points, so that each pixel can be distributed to the neighbor of the original clustering center;
●, the above operations are performed iteratively so that all pixels converge on a particular cluster center;
3) the nearest neighbors match. Performing nearest neighbor matching on the super-pixels of the regions of interest and the super-pixels of the template, and finding the nearest neighbor of the super-pixels in the super-pixel set of the template for the super-pixels in each region of interest:
NN(q,PROI)＝argminp∈Pd(pA,qA)
wherein d (-) represents the pixel distance, PROIRepresenting a superpixel set resulting from the segmentation of the region of interest, Q representing a superpixel set resulting from the segmentation of the template region, and pA、qAThe characteristics of the superpixels in set P, Q are respectively represented.
4) A diversity measure. The logarithm between the template point and the target point obtained by nearest neighbor matching can be used to measure the similarity between the two, so when defining the template superpixel point set as
And the target superpixel point set is
In time, the diversity metric method can be formulated as:
w(qj)＝NNW(qj,P)-k(NN(qj,P))
wherein the content of the first and second substances,
is a regularization factor, k (p)i) The fraction of the nearest neighbor corresponding to the target superpixel found in the template superpixel set is determined.
5) And calculating the weight of the particles. Uniformly substituting the similarity measurement results into a Gaussian distribution (taking the measured value as a mean value and the measured noise as a variance) for calculation, wherein the obtained result is the weight of the particle state
And 4, estimating the state of the particles. Normalizing the weight of the particles, and performing weighted summation by combining the particle states, wherein the summation result is the estimated target state:
and 5, resampling the particles. The weight of the accumulated particles
And selecting a random number m-R (0,1) in accordance with uniform distribution, and determining an order of arrival using the random number
Particles of established state
Adding it into the particle set and initializing its state weight as 1/N;
the effects of the present invention can be further illustrated by the following experiments:
1. conditions of the experiment
The simulation platform adopted in the experiment is matlab R2017a, and the adopted simulation environment has conditions of Intel (R) core (TM) i7-6700(3.4GHZ) CPU, 32.0GB memory, 64-bit operating system and the like.
2. Contents and results of the experiments
The experimental data adopted in fig. 3a, 3b, 4a and 4b are all from videos aerial by unmanned aerial vehicles under field conditions. In this video, the tracker needs to track two jogging girls, and needs to overcome visual challenges including out-of-plane rotation, object occlusion, motion blur, and perspective transformation. It can be seen that the diversity metric based particle filter tracker rises faster in distance accuracy than the KCF tracker and falls slower in overlay accuracy than the KCF tracker in tracking brown and white girls in video. Therefore, the method provided by the invention is better than the existing KCF tracking method in terms of tracking effect.
Fig. 5 measures the performance of the tracking method from the viewpoint of processing each frame time by the tracking algorithm. As can be seen from the graphs, as the number of sampled particles increases, the graph of the LOT algorithm diverges, while the graph of the diversity metric particle filter algorithm converges. This means that the diversity metric particle filter algorithm can improve the effect of the algorithm by increasing the number of the sampling particles without increasing the time cost, so that the method has higher popularization and application value than the traditional particle filter algorithm.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.
Claims (5)
1. A pedestrian robust tracking method for field search and rescue of unmanned aerial vehicles is characterized by comprising the following steps:
step 1, initializing a video image, namely calibrating a tracking target in a given initial frame, and performing random sampling around the tracking target to form an initialized particle set, specifically:
step 1.1, read in initialization frame I0Acquiring a proper target from the target as a tracking template;
step 1.2, initializing and defining the state space of the target by utilizing the position and size information of the template to obtain the initial state X of the target0；
Step 1.3, according to the initial state X of the target0Randomly generating N particles to form an initialized particle set
Wherein k-1 represents the time k-1,
obeying a mean value of X0The variance is a Gaussian distribution of phi, and
then the weight value is normalized by the particles and is initialized to 1/N;
step 2, particle state propagation, namely propagating each particle in the particle set according to a state propagation model of the system, so that the propagated particles obey a known importance probability distribution, including:
step 2.1, region-of-interest extraction, namely, each particle in the single view particle set is not considered any more, but the region formed by the particles is considered as a region-of-interest (ROI), and the region-of-interest is treated as a whole, specifically, the particle set at the time k is acquired
Fusing the particles to obtain a region of interest (ROI) at the K moment;
step 2.2, performing superpixel segmentation, namely extracting pixels of the region of interest and the template, and expressing the characteristics in a mode of replacing the pixels with the superpixels;
step 2.3, nearest neighbor matching; performing nearest neighbor matching on the super-pixels of the regions of interest and the super-pixels of the template, and finding the nearest neighbor of the super-pixels in the super-pixel set of the template for the super-pixels in each region of interest;
step 2.4, diversity measurement, namely utilizing the idea of nearest neighbor matching to carry out similarity measurement so as to realize the updating of the particle weight;
step 2.5, calculating the weight of the particles; uniformly substituting the similarity measurement results into a Gaussian distribution for calculation, wherein the obtained result is the weight of the particle state
And 3, estimating the particle state, namely performing normalization operation on the weight of the particles, and performing weighted summation by combining the states of the particles to obtain the current tracking targetThe state of the frame, specifically, the weight of the normalized particle, is weighted and summed by combining the particle states, and the result of the summation is the estimated target state:
step 4, resampling the particles, namely screening the particles from the angle of probability statistical distribution, so as to inhibit the degradation of the state weight of the particles, specifically: the weight of the accumulated particles
And selecting a random number m-R (0,1) in accordance with uniform distribution, and determining an order of arrival using the random number
Particles of established state
It is added into the particle set and initialized to have a state weight of 1/N.
2. The robust pedestrian tracking method for field search and rescue of unmanned aerial vehicles according to claim 1, wherein the step 2 is realized by the method comprising the following steps: defining an importance probability distribution satisfies the following constraints:
q(xk|x0:k-1,y1:k)＝q(xk|xk-1,yk)
wherein xkAnd ykRespectively representing the state value and the measured value of the particle at the time k, q (-) represents the probability distribution of importance, and
it represents the state noise of the system (which may be represented bySolving a system state transition equation); it can be seen that the propagation process of the particle can be predicted according to the state transition equation, and the propagated particle can be obtained through the markov process; thus, the sub-state propagation will be done in two steps:
step 2.1, obtaining the particle set at the k-1 moment
And propagating the state of the particles to obtain a corresponding prediction result:
wherein v isk-1Obeying a Gaussian distribution with the mean value of 0 and the variance of phi as state noise;
step 2.2, the particles obtained by the transmission are combined to obtain a state set of the system at the time k
Wherein the state of the particles is represented by
The propagation is obtained and the weight of the particle is calculated from
The transfer is obtained.
3. The robust pedestrian tracking method for field search and rescue of unmanned aerial vehicles according to claim 1, wherein the step 2.2 is realized by the method comprising the following steps:
step A, dividing an image into grids, and determining the center of each grid as an initialization clustering center;
b, distributing the pixel points in the grid to the nearest cluster center by using a k-means method;
c, adjusting the clustering center according to the characteristics of the pixel points, so that each pixel can be distributed to the neighbor of the original clustering center;
and step D, iteratively performing the operations from the step A to the step C, so that all pixels can converge on a specific clustering center.
4. The robust pedestrian tracking method for field search and rescue of unmanned aerial vehicles according to claim 1, wherein the step 2.3 is realized by the method comprising the following steps: performing nearest neighbor matching on the super-pixels of the regions of interest and the super-pixels of the template, and finding the nearest neighbor of the super-pixels in the super-pixel set of the template for the super-pixels in each region of interest:
NN(q,PROI)＝argminp∈Pd(pA,qA)
wherein d (-) represents the pixel distance, PROIRepresenting a superpixel set resulting from the segmentation of the region of interest, Q representing a superpixel set resulting from the segmentation of the template region, and pA、qAThe characteristics of the superpixels in set P, Q are respectively represented.
5. The robust pedestrian tracking method for field search and rescue of unmanned aerial vehicles according to claim 1, wherein the step 2.4 is realized by the method comprising the following steps: defining a template superpixel point set as
And the target superpixel point set is
The diversity metric method is formulated as:
w(qj)＝NNW(qj,P)-k(NN(qj,P))
wherein the content of the first and second substances,
is a regularization factor, k (p)i) The fraction of the nearest neighbor corresponding to the target superpixel found in the template superpixel set is determined.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710835495.3A CN107563342B (en) | 2017-09-15 | 2017-09-15 | Pedestrian robust tracking method for field search and rescue of unmanned aerial vehicle |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710835495.3A CN107563342B (en) | 2017-09-15 | 2017-09-15 | Pedestrian robust tracking method for field search and rescue of unmanned aerial vehicle |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN107563342A CN107563342A (en) | 2018-01-09 |
| CN107563342B true CN107563342B (en) | 2020-03-10 |
Family
ID=60980083
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710835495.3A Expired - Fee Related CN107563342B (en) | 2017-09-15 | 2017-09-15 | Pedestrian robust tracking method for field search and rescue of unmanned aerial vehicle |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107563342B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110472588B (en) * | 2019-08-19 | 2020-11-24 | 上海眼控科技股份有限公司 | Anchor point frame determining method and device, computer equipment and storage medium |
| CN112288777A (en) * | 2020-12-16 | 2021-01-29 | 西安长地空天科技有限公司 | Method for tracking laser breakpoint by using particle filtering algorithm |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102034245A (en) * | 2010-12-17 | 2011-04-27 | 浙江大学 | Method for tracking landmark on unmanned helicopter platform |
| CN102831409A (en) * | 2012-08-30 | 2012-12-19 | 苏州大学 | Method and system for automatically tracking moving pedestrian video based on particle filtering |
| CN105809206A (en) * | 2014-12-30 | 2016-07-27 | 江苏慧眼数据科技股份有限公司 | Pedestrian tracking method |
-
2017
- 2017-09-15 CN CN201710835495.3A patent/CN107563342B/en not_active Expired - Fee Related
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102034245A (en) * | 2010-12-17 | 2011-04-27 | 浙江大学 | Method for tracking landmark on unmanned helicopter platform |
| CN102831409A (en) * | 2012-08-30 | 2012-12-19 | 苏州大学 | Method and system for automatically tracking moving pedestrian video based on particle filtering |
| CN105809206A (en) * | 2014-12-30 | 2016-07-27 | 江苏慧眼数据科技股份有限公司 | Pedestrian tracking method |
Non-Patent Citations (2)
| Title |
|---|
| Pedestrian Tracking in Infrared Image;Jianfu Li etal.;《2009 Second Asia-Pacific Conference on Computational Intelligence and Industrial Applications》;20091231;全文 * |
| 邻域迭代重采样粒子滤波的纯方位目标跟踪;王向前等;《计算机工程与应用》;20160708;第53卷(第13期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN107563342A (en) | 2018-01-09 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Dornaika et al. | Building detection from orthophotos using a machine learning approach: An empirical study on image segmentation and descriptors | |
| Shen et al. | Interactive segmentation using constrained Laplacian optimization | |
| Li et al. | Adaptive deep convolutional neural networks for scene-specific object detection | |
| CN102741884A (en) | Mobile body detection device and mobile body detection method | |
| CN109242019B (en) | Rapid detection and tracking method for optical small target on water surface | |
| Grudic et al. | Outdoor Path Labeling Using Polynomial Mahalanobis Distance. | |
| CN103985143A (en) | Discriminative online target tracking method based on videos in dictionary learning | |
| Li et al. | Unsupervised road extraction via a Gaussian mixture model with object-based features | |
| CN107563342B (en) | Pedestrian robust tracking method for field search and rescue of unmanned aerial vehicle | |
| Bourouis et al. | Color object segmentation and tracking using flexible statistical model and level-set | |
| Alsanad et al. | Real-time fuel truck detection algorithm based on deep convolutional neural network | |
| Wang et al. | HypLiLoc: Towards effective LiDAR pose regression with hyperbolic fusion | |
| Ren et al. | Research on infrared small target segmentation algorithm based on improved mask R-CNN | |
| US8478032B2 (en) | Segmenting an image | |
| Mei et al. | Geodesic‐based probability propagation for efficient optical flow | |
| Chester et al. | Machine learning for image classification and clustering using a universal distance measure | |
| Cheng et al. | Parallel particle filter for tracking road centrelines from high-resolution SAR images using detected road junctions as initial seed points | |
| CN114677508A (en) | Point cloud instance semantic segmentation method based on dynamic filtering and point-by-point correlation | |
| Mei et al. | Fast template matching in multi-modal image under pixel distribution mapping | |
| Yan et al. | Repeatable adaptive keypoint detection via self-supervised learning | |
| Haltakov et al. | Geodesic pixel neighborhoods for 2D and 3D scene understanding | |
| Zhou et al. | Gaussian-weighted Jensen–Shannon divergence as a robust fitness function for multi-model fitting | |
| Shimosaka et al. | Counting pedestrians in crowded scenes with efficient sparse learning | |
| Zou et al. | Scene flow for 3D laser scanner and camera system | |
| Zhan et al. | Hyperspectral image classification using multilayer superpixel graph and loopy belief propagation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20200310 Termination date: 20210915 |