CN104778443A - Natural road sign recognition based on real-time local characteristic description - Google Patents
Natural road sign recognition based on real-time local characteristic description Download PDFInfo
- Publication number
- CN104778443A CN104778443A CN201510098660.2A CN201510098660A CN104778443A CN 104778443 A CN104778443 A CN 104778443A CN 201510098660 A CN201510098660 A CN 201510098660A CN 104778443 A CN104778443 A CN 104778443A
- Authority
- CN
- China
- Prior art keywords
- algorithm
- real
- point
- time
- characteristic
- 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.)
- Pending
Links
Landscapes
- Image Analysis (AREA)
Abstract
The invention adopts a characteristic extracting technology of a local characteristic matching technology and performs characteristic matching by using a method combining characteristic extraction and binary characteristic description based on a curvature in accordance with road sign matching scene application requirements of outdoor scenes. An algorithm utilizes reasonable characteristic of characteristic distribution based on a curvature algorithm and solves the problem that the distribution of characteristics is uneven, the defect that the real-time capability of other algorithms is weaker is overcome through the binary description algorithm, and finally the validity and the real-time capability of the algorithm are contrasted through experiments. Results indicate that the algorithm can ensure the real-time capability, at the same time, effective even characteristic points are extracted, and the algorithm has better practicality and robustness.
Description
technical field:
Patent of the present invention relates to the natural landmark identification based on real-time local feature description.
background technology:
Location is one of the most basic function of mobile robot; the problem that the present invention mainly solves is exactly orientation problem; Position Research achievement is the earliest mainly by the internal sensor of robot; as the facility such as code table, inertia apparatus realizes location; and the deviation accumulation caused due to the reason such as to skid, ground is rugged and rough usually can cause positioning result out of true, unsuitable long-time navigation separately.And if located by GPS or China triones navigation system location, having the local positioning system of the poor signals such as occlusion area to lose efficacy, therefore, people start to gradually adopt external sensor to carry out auxiliary positioning.Such as infrared and vision sensor.
Wherein, environmental information that vision sensor enriches because of it receives much concern as vision, texture, shape etc., and known by theory on computer vision, just accurately can identify target, and judge self-position by vision system.In current many existing methods, mostly have employed manually-set road sign, then in robot moving process, coupling is carried out realize location to the artificial landmark of setting in advance.But along with the continuous progress of research work, people have forwarded the application scenarios of mobile robot to outdoor gradually indoor, now under many particular cases, manually-set road sign is unpractical, so natural landmark just becomes the prefered method under condition out of doors.So-called physical feature be exactly existing in finger ring border, unartificial setting, can in order to identify the feature object of varying environment scene.And relatively indoor most structurized environment, outdoor complicated unstructured moving grids brings sizable challenge to the location of mobile robot and navigation, meanwhile, relative to indoor, the illumination condition of outdoor constantly change and climatic environment are also a major challenge in research.
The natural landmark of view-based access control model needs to extract its invariant point, utilizes the method for Feature Points Matching afterwards to judge and identifies road sign, thus realizing the location of mobile robot.With directly use the method for sensor different, first original sampled data can be converted to corresponding local feature by the localization method of feature based, saves abundant and accurate environmental information when only needing a small amount of storage space.These class methods have good robustness usually, can meet the application of relative complex.But sometimes in urban transportation, most of unique point concentrates on the trees of road both sides, and trees of the same race are also more similar, and this all brings larger impact to the characteristic matching in later stage and robot localization.Therefore, how to reject the impact that trees are extracted natural landmark, unique point is distributed in landmark thing as far as possible uniformly and goes, then become natural landmark extract in a large complicated and difficult task, extracting reliable and abundant feature is key in feature location.
Current existing local feature matching algorithm mainly contains Harris, SIFT, SURF, ORB etc., SIFT(Scale-invariant feature transform) algorithm delivered in 1999 by David Lowe, and within 2004 years, improve and sum up.Extensive concern is subject to because it has good dimension rotation unchangeability, but comparatively slow in arithmetic speed due to it, have impact on its utilization in application in real time.Rear Herbert Bay proposes the SURF algorithm of improvement for SIFT algorithm, and speed improves a number magnitude, but this speed still not in some in real time application.Until willowgarage in 2011 proposes ORB algorithm, the application of most of real-time can be met.But the angle point that above-mentioned three kinds of methods are extracted, out of doors in application, unique point distribution is even all not.The algorithm based on curvature that the present invention proposes improvement realizes more uniform extract minutiae.
summary of the invention:
1 feature point extraction based on curvature improved
Feature point extraction method based on curvature calculates simple, and the unique point extracted has rotational invariance, also all have good robustness, and unique point is evenly distributed to noise and brightness change, and the more important thing is can the unique point of Extraction and determination.The method judges whether it is angle point mainly through grey scale change, and when in image, certain a bit all can produce a large amount of changes of gray scale along the minor shifts of any direction, then this point is considered to angle point.If
for the change of horizontal direction gray scale,
for the grey scale change of vertical direction, so angle point is exactly
,
change all very large point, edge is then
,
in only have one to change greatly point.If
for with
centered by a slide block, when it slides [u, v] in any direction, produce grey scale change
computing method as follows:
（1）
Then go out the partial derivative of single order to n rank according to taylor series computation, finally can obtain a Matrix Formula:
（2）
Then according to the eigenwert of matrix
determine whether angle point, but it is as follows generally to calculate angle point response in practical operation:
（3）
Wherein k is coefficient value, generally gets between 0.04 to 0.06, determines whether angle point by R, if the maximum value of the corresponding neighborhood of R is exactly angle point position.
The feature point extraction algorithm based on curvature originally, judges whether a point is calculating according to judgement of unique point
whether maximum in neighborhood, if maximum, it is unique point, and here because the uncertainty of a point depends primarily on that less eigenwert, eigenwert according to determinant determines whether the words of unique point, a less eigenwert numerical value is only needed to be greater than the threshold value pre-set, then this point just can be considered to a strong angle point, namely now:
（4）
If
be greater than prior given threshold value and be strong angle point, because this algorithm is to finding that the requirement of eigenwert is lower, can be drawn into some is not obvious especially unique point, if when in practical application, characteristics of needs point is uniformly distributed, the computing neighborhood of suitable this algorithm of raising, can obtain the more uniform angle point distribution plan of other algorithm distributions relatively.
2 binary features improved describe algorithm
Binary-coded character string descriptor, other algorithms decrease the dimension of descriptor relatively, thus decrease the time describing and mate.First this algorithm carries out gaussian filtering to reduce the impact of noise to image, after select a localized mass p in the picture, size is SxS pixel, defines one
test as follows:
（5）
Wherein u, v are the two-dimensional coordinate pair of shape as (x, y), and p (u) p (v) is the brightness of point.Then again by several
the result composition bit string of test:
（6）
Here n can get 128,256,512 etc., and the present invention gets 512, and it needs the bytes of storage space of 512/8=64.How effective selected characteristic point, on very large on ensuing utilization impact, the experiment proved that, adopts and obey
gaussian distribution time, proper vector will have good resolution, simultaneously adopt Hamming distance from replacement traditional Euclidean distance judge whether coupling, only need step-by-step to carry out xor operation during coupling, substantially increase the time of characteristic matching.
BRIEF computing is simple and shared internal memory is less, compares to be applicable to some and to apply in real time, but this algorithm also exists some shortcomings equally: easily affected by noise; Do not consider characteristic direction, do not possess rotational invariance; Do not possess scale invariability.
For allowing algorithm, there is rotational invariance and rotate to reduce mobile robot's picture in dither process the impact that matching result is caused, calculated the direction of each key point by centroid algorithm, be shown below:
（7）
Obtaining unique point direction is:
（8）
Wherein
for the coordinate of point of interest being detected, afterwards by obtain
with
footage number quantized.In position
on to any n binary features collection, definition a 2*n matrix:
（9）
With obtaining before
correct
obtain
, wherein
（10）
Descriptor is extracted again in direction according to obtaining
just there is rotational invariance.
Accompanying drawing illustrates:
Fig. 1 is based on the design sketch of the feature point extraction of curvature.Figure (a) is maximum value suppression neighborhood
time the design sketch of the feature point extraction based on curvature; Figure (b) is maximum value suppression neighborhood
time the design sketch of the feature point extraction based on curvature.According to formula (4) if
be greater than prior given threshold value and be strong angle point, because this algorithm is to finding that the requirement of eigenwert is lower, can be drawn into some is not obvious especially unique point, if when in practical application, characteristics of needs point is uniformly distributed, the computing neighborhood of suitable this algorithm of raising, can obtain the more uniform angle point distribution plan of other algorithm distributions relatively.
Fig. 2 is based on the uniform characteristics extraction effect figure of the curvature improved.A () figure is that the maximum value improved suppresses neighborhood
time the uniform characteristics extraction effect figure based on curvature, (b) figure be improve maximum value suppress neighborhood
time the uniform characteristics extraction effect figure based on curvature, can find to expand as when maximum value inhibition zone
time, feature is also quite even with distribution.
By under win7 system, the natural landmark utilizing vs2010 and opencv2.4.10 to achieve even Harris-brief is extracted.Experimental Hardware is the A6-3420M APU1.5GHz of AMD tetra-core, internal memory 4G, video card HD 7470M 1G.Few owing to realizing feature point pairs needed for localization for Mobile Robot, therefore the present invention to count to the feature that each sub-picture detects to all algorithms, to arrange the upper limit be 200, use picture to adopt close frame screenshotss gained figure in the video of one section of running car.
Fig. 3 is each algorithm experimental effect.Figure (a) is Fast Corner algorithm effect figure, and figure (b) is SIFT extraction algorithm design sketch, and figure (c) is SURF extraction algorithm design sketch, and figure (d) is for former based on Curvature Matching algorithm effect figure, and figure (e) improves the design sketch based on curvature algorithm.
Fig. 4 is the correlation data of a few class algorithm.
Practical function:
1) from effect, the unique point that SIFT and SURF obtains distributes more even equally, but compare with other algorithms in speed, slow a lot, be not suitable for real-time application, and the unique point distribution that fastest Fast Corner Detection algorithm detects is comparatively concentrated, the impact of the things such as trees on characteristic matching cannot be reduced.
2) based on curvature algorithm relatively other algorithms for this application undoubtedly advantageously, the Feature Extraction Method based on curvature of improvement, the processing time is compared with former algorithm, should reduce characteristic standard, simplify calculating process, and the time is short all the better.Correct matching rate aspect is more or less the same, and for the aspect that is evenly distributed that the present invention mainly studies, can find relatively former algorithm, and distribution is more evenly a lot, decreases the impact that the natural landmark such as trees cause characteristic matching.Therefore the present invention proposes under algorithm can meet open-air conditions, the task that effective extract minutiae carrying out mates, and there is good real-time and robustness.
Embodiment:
1. based on the natural landmark identification of real-time local feature description, it is characterized in that: adopt the feature point extraction based on curvature improved, feature point extraction method based on curvature calculates simple, the unique point extracted has rotational invariance, also all there is good robustness to noise and brightness change, and unique point is evenly distributed, the more important thing is can the unique point of Extraction and determination.The method judges whether it is angle point mainly through grey scale change, and when in image, certain a bit all can produce a large amount of changes of gray scale along the minor shifts of any direction, then this point is considered to angle point.
2. based on the natural landmark identification of real-time local feature description, it is characterized in that: adopt the binary features improved to describe algorithm, binary-coded character string descriptor, other algorithms decrease the dimension of descriptor relatively, thus decrease the time describing and mate.First this algorithm carries out gaussian filtering to reduce the impact of noise to image, after select a localized mass p in the picture, size is SxS pixel.
3. based on the natural landmark identification of real-time local feature description, it is characterized in that: other algorithms relative of the algorithm based on curvature for this application undoubtedly advantageously, the Feature Extraction Method based on curvature improved, processing time is compared with former algorithm, characteristic standard should be reduced, simplify calculating process, the time is short all the better.Correct matching rate aspect is more or less the same, and for the aspect that is evenly distributed that the present invention mainly studies, can find relatively former algorithm, and distribution is more evenly a lot, decreases the impact that the natural landmark such as trees cause characteristic matching.Therefore the present invention proposes under algorithm can meet open-air conditions, the task that effective extract minutiae carrying out mates, and there is good real-time and robustness.
Claims (3)
1. based on the natural landmark identification of real-time local feature description, it is characterized in that: adopt the feature point extraction based on curvature improved, feature point extraction method based on curvature calculates simple, the unique point extracted has rotational invariance, also all there is good robustness to noise and brightness change, and unique point is evenly distributed, the more important thing is can the unique point of Extraction and determination; The method judges whether it is angle point mainly through grey scale change, and when in image, certain a bit all can produce a large amount of changes of gray scale along the minor shifts of any direction, then this point is considered to angle point.
2. based on the natural landmark identification of real-time local feature description, it is characterized in that: adopt the binary features improved to describe algorithm, binary-coded character string descriptor, other algorithms decrease the dimension of descriptor relatively, thus decrease the time describing and mate; First this algorithm carries out gaussian filtering to reduce the impact of noise to image, after select a localized mass p in the picture, size is SxS pixel.
3. based on the natural landmark identification of real-time local feature description, it is characterized in that: other algorithms relative of the algorithm based on curvature for this application undoubtedly advantageously, the Feature Extraction Method based on curvature improved, processing time is compared with former algorithm, characteristic standard should be reduced, simplify calculating process, the time is short all the better; Correct matching rate aspect is more or less the same, and for the aspect that is evenly distributed that the present invention mainly studies, can find relatively former algorithm, and distribution is more evenly a lot, decreases the impact that the natural landmark such as trees cause characteristic matching; Therefore the present invention proposes under algorithm can meet open-air conditions, the task that effective extract minutiae carrying out mates, and there is good real-time and robustness.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510098660.2A CN104778443A (en) | 2015-03-06 | 2015-03-06 | Natural road sign recognition based on real-time local characteristic description |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510098660.2A CN104778443A (en) | 2015-03-06 | 2015-03-06 | Natural road sign recognition based on real-time local characteristic description |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN104778443A true CN104778443A (en) | 2015-07-15 |
Family
ID=53619896
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510098660.2A Pending CN104778443A (en) | 2015-03-06 | 2015-03-06 | Natural road sign recognition based on real-time local characteristic description |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104778443A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106651878A (en) * | 2016-12-21 | 2017-05-10 | 福建师范大学 | Method for extracting straight line from local invariant feature points |
| CN108376408A (en) * | 2018-01-30 | 2018-08-07 | 清华大学深圳研究生院 | A kind of three dimensional point cloud based on curvature feature quickly weights method for registering |
| CN117119253A (en) * | 2023-06-28 | 2023-11-24 | 三峡科技有限责任公司 | High-quality video frame extraction method for target object |
-
2015
- 2015-03-06 CN CN201510098660.2A patent/CN104778443A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106651878A (en) * | 2016-12-21 | 2017-05-10 | 福建师范大学 | Method for extracting straight line from local invariant feature points |
| CN106651878B (en) * | 2016-12-21 | 2019-06-11 | 福建师范大学 | A method of for extracting straight line from local invariant feature point |
| CN108376408A (en) * | 2018-01-30 | 2018-08-07 | 清华大学深圳研究生院 | A kind of three dimensional point cloud based on curvature feature quickly weights method for registering |
| CN117119253A (en) * | 2023-06-28 | 2023-11-24 | 三峡科技有限责任公司 | High-quality video frame extraction method for target object |
| CN117119253B (en) * | 2023-06-28 | 2024-05-07 | 三峡科技有限责任公司 | High-quality video frame extraction method for target object |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2016034059A1 (en) | Target object tracking method based on color-structure features | |
| CN104036524A (en) | Fast target tracking method with improved SIFT algorithm | |
| CN105528794A (en) | Moving object detection method based on Gaussian mixture model and superpixel segmentation | |
| CN103310194A (en) | Method for detecting head and shoulders of pedestrian in video based on overhead pixel gradient direction | |
| CN107066969A (en) | A kind of face identification method | |
| CN105279769A (en) | Hierarchical particle filtering tracking method combined with multiple features | |
| CN104794737A (en) | Depth-information-aided particle filter tracking method | |
| CN103824305A (en) | Improved Meanshift target tracking method | |
| CN103106409A (en) | Composite character extraction method aiming at head shoulder detection | |
| CN104778443A (en) | Natural road sign recognition based on real-time local characteristic description | |
| CN102902973B (en) | A kind of extracting method with the characteristics of image of rotational invariance | |
| CN110738098A (en) | target identification positioning and locking tracking method | |
| Edoardo Maresca et al. | The Matrioska tracking algorithm on LTDT2014 dataset | |
| Duan et al. | An approach to dynamic hand gesture modeling and real-time extraction | |
| Choudri et al. | Robust background model for pixel based people counting using a single uncalibrated camera | |
| Zhou et al. | SURF feature detection method used in object tracking | |
| Jin et al. | A method for traffic signs detection, tracking and recognition | |
| Yang et al. | Cascaded superpixel pedestrian object segmentation algorithm | |
| Park et al. | Global localization using low-frequency image-based descriptor and range data-based validation | |
| CN107273850B (en) | Autonomous following method based on mobile robot | |
| Wang et al. | Combining color histogram and ORB features for robust visual tracking | |
| Jianbo et al. | Moving target tracking algorithm based on scale invariant optical flow method | |
| Li et al. | A mean shift tracking algorithm based on the current statistical model | |
| Li et al. | Real-time depth-based segmentation and tracking of multiple objects | |
| Xu et al. | A VSS Algorithm Based on Multiple Features for Object Tracking. |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20150715 |