Technical Field
The present invention relates to unmanned technical field, in particular to a based on ultrasonic sensor scene re-recognition algorithm.
Background Art
Unmanned technology in the covers five major module; positioning, constructs the chart, perception, planning and control. In the closed application scene (such as car, within the complex, such as in a cell) of the unmanned solution in one of the most important function of the is, according to the sensing module after the processing of the sensor information from the scene map to identify the unmanned vehicle present in the environment, the function realization depends on three aspects; accurate map information, precise and reliable sensor and sensor information and the map matching comparison algorithms scene re-identification algorithm. In recent years to include the high-precision laser radar, long distance millimeter wave radar, short distance ultrasonic radar and camera sensor including the development of the above three aspects of the front two provided in the support. The present reliable scene re-identification algorithm based on the majority of the real-time laser radar of the real-time data with the prebuilt map comparison, because of the high laser radar shall cost and the computer to calculate the capacity constraints, using the algorithm of the unmanned vehicle solution cost high. There is at present no suitable for such as ultrasonic sensor device this low-cost sensor scene re-recognition algorithm.
Content of the invention
One of the purposes of the present invention is to solve the at least the above-mentioned problems, and to provide at least the rear and the advantages of the note.
The invention has another purpose is to provide a based on ultrasonic sensor scene re-recognition algorithm, set up and the ultrasonic sensor matching scene re-recognition algorithm, greatly reduces the scene re-identification of the cost and power consumption.
In order to achieve these purposes according to the present invention and other advantages, provides a based on ultrasonic sensor scene re-recognition algorithm, including:
Step 1, pre-stored in the map database in the scene, the scene in the extracted identification information of the map and the identification information stored in the database, the identification information includes the scene map in the curved surface of the scene or object, boundary line, and the semantic feature point of the corner spot.
Step 2, in the step 1 in the scene or object and the boundary line of the curved surface fitting, in order to get the scene or object information of the fitting, and the fitting information storage to the database.
Step 3, in the course of driving, in accordance with time-sequential, the unmanned vehicle of any one of the ultrasonic sensor real-time dynamic access to its relative to the detection range of the obstacle of the plurality of real-time position information, the real-time dynamically obtaining real-time position information of the plurality of coordinate transformation of sequentially a plurality of coordinate information.
Step 4, according to the time sequence, are sequentially connected in series real-time dynamic acquisition of any one of the ultrasonic sensor of the plurality of coordinate information in order to obtain the ultrasonic sensor changes over time of the track, and so on, to obtain the unmanned vehicle of the plurality of ultrasonic transducer a plurality of tracks, a plurality of orbits to make up for the unmanned vehicle of the running track.
Step 5, through the extraction step 4 in [...] track in the plurality of track and curve corner spot, the corner point mark and the curve fitting, in order to be track fitting information.
Step 6, through the step 5 in the track fitting information with the step 3 to the database stored in said extracting information and scene fitting information match the contrast, in order to optimize the unmanned vehicle to the ultrasonic sensor to sense the environment most matching scene local region; wherein said match the contrast is abstracted to curve and manifold of the optimization of the difference metric.
Preferably, the step 1 pre-stored in the map of said scene by laser radar or millimeter wave radar to obtain.
Preferably, the step 1 in the corner point of the Harris operator identification; the identification of the antithetical feature points according to the computing power restrictions and precision requirement of the SIFT operator, accelerate the sound characteristic, or through a HOG a pixel with its surrounding field of a sufficient number of pixels in the difference between the larger mode, determining said pixel may be the corner of the FAST operator; the boundary line of using Canny algorithm.
Preferably, step 5 in the orbit fitting information acquisition process are as follows:
Using the Harris operator of said corner point identification.
The gradient descent, Gaussian process or EM algorithm to the curve fitting.
Preferably, step 2 in fitting the edge boundaries of the [...]; the curved surface and the curved surface on the curve fitting using the gradient descent, Gaussian process or EM algorithm.
Preferably, step 6 in the metric can to said curved surface after slicing the use of L2 distance, or based on the probability distribution of the dispersion of the KL divergence or JS.
Preferably, step 6 in the optimizing algorithm is the gradient drop law.
The invention comprises at least the following advantages:
The invention obtains scene map, in this scene map can be used for the map matching contrast on the scene or the curved surface of the inner Figures, boundary line, and corner spot semantic feature point extracting identification information such as out in the storage to the database, the backup transfer is easy to use, and the extracted boundary line and the connecting line of the curved surface to point, and the connecting line of the graphics or data information stored in the database, wherein the database is the memory is relatively large, the query efficiency higher efficient database; through the unmanned vehicle in real-time reading of the ultrasonic sensor of the body around the location of the obstacle information, the determination of the position of the obstacle information needed without taking into account the person in the course of driving the vehicle and the running track of the position of the ultrasonic sensor relative to the vehicle body, the location of the obstacle information is converted to a coordinate form, convenient in prebuilt scene positioning it on the map, so as to establish a running track; by extracting the identification unmanned vehicle in the running track of the corner point label, and the curve, is established for the map matching comparison information, in the database storage of the scene map related to the curved surface of the fitting and after the feature point information, matching comparison, feature point, curves and surfaces matching comparison can be abstract mathematical optimization curve and manifold of the difference metric, thus will match the comparison process-specific to can be seen on the optimization algorithm, in order to be with the ultrasonic sensor sensing the most matching local region, is used as a real-time laser radar of the real-time data comparison with the prebuilt map by the high cost of the need to bear, through the scene re-identification algorithm, has realized the use of the ultrasonic sensor and the purpose of the scene re-identification, greatly reducing realizes the scene re-identification function of the cost and power consumption.
Other advantages of the present invention, objectives and features will reflect through the lower part of the note, segment will also be through to the study and practice of this invention is in the field of the technical understood.
Description of drawings
Figure 1 is the flow chart of the invention based on ultrasonic sensor scene re-recognition algorithm.
Mode of execution
The Figure below to the further detailed description of this invention, in order to make the technical personnel in the field specification can be on the basis of the implementation of the reference characters.
It should be understood, used herein such as "has", "comprising" and "including" terminology does not exclude one or a plurality of other components or a combination thereof the presence or added.
As shown in Figure 1, the present invention provides a based on ultrasonic sensor scene re-recognition algorithm, including:
Step 1, by acquiring the scene map, extracting identification map of the scene or object in the scene of the curved surface, boundary line, and corner spot semantic characteristic point, in order to be identification information.
Step 2, in the step 1 in the surface fitting states the boundary line and, in order to get the scene fitting information.
Step 3, the step 1 and the step of extracting information 2 fitting of said scene information storage to the database.
Step 4, in the absence of people and vehicles through the arrangement of the ultrasonic sensor on the real-time reads the [...] during the running of the vehicle, location of the obstacle information, the location information to the coordinate transformation, in order to obtain said [...] the running track of the vehicle.
Step 5, through the extraction step 4 in the corner spot in a track [...] and curve, the corner point mark and the curve fitting, in order to be track fitting information.
Step 6, through the step 5 in the track fitting information with the step 3 to the database stored in said extracting information and scene fitting information match the contrast, in order to optimize the unmanned vehicle to the ultrasonic sensor to sense the environment most matching scene local region; wherein said match the contrast is abstracted to curve and manifold of the optimization of the difference metric.
In the above-mentioned scheme in, through the prebuilt scene map, in this scene map can be used for the map matching contrast on the scene or the curved surface of the inner Figures, boundary line, and corner spot semantic feature point extracting identification information such as out in the storage to the database, the backup transfer is easy to use, and the extracted boundary line and the connecting line of the curved surface to point, and the connecting line of the graphics or data information stored in the database, wherein the database is the memory is relatively large, the query efficiency higher efficient database; through the unmanned vehicle in real-time reading of the ultrasonic sensor of the body around the location of the obstacle information, the determination of the position of the obstacle information needed without taking into account the person in the course of driving the vehicle and the running track of the position of the ultrasonic sensor relative to the vehicle body, the location of the obstacle information is converted to a coordinate form, convenient in prebuilt scene positioning it on the map, so as to establish a running track; by extracting the identification unmanned vehicle in the running track of the corner point label, and the curve, is established for the map matching comparison information, in the database storage of the scene map related to the curved surface of the fitting and after the feature point information, matching comparison, feature point, curves and surfaces matching comparison can be abstract mathematical optimization curve and manifold of the difference metric, thus will match the comparison process-specific to can be seen on the optimization algorithm, in order to be with the ultrasonic sensor sensing the most matching local region, is used as a real-time laser radar of the real-time data comparison with the prebuilt map by the high cost of the need to bear, through the scene re-identification algorithm, has realized the use of the ultrasonic sensor and the purpose of the scene re-identification, greatly reducing realizes the scene re-identification function of the cost and power consumption.
In one preferred embodiment, step 1 pre-stored in the map of the scene by laser radar or millimeter wave radar to obtain.
In the above-mentioned scheme in, through the laser radar to to a target detection signal, then the received reflected from the target back to compare the signal with the transmitted signals, after appropriate processing, to obtain the target relevant information, such as target distance, orientation, height, speed, attitude, even shape and other parameters, so that the unmanned in the region of the scene map prebuilt more comprehensive, accurate and efficient; millimeter wave radar is work in millimetric wave band detecting radar, millimeter wave radar can distinguish the goal of identification is small, but also can simultaneously identify a plurality of target function, can also be used for the map for the establishment of the scene.
In one preferred embodiment, step 1 in the corner point of the Harris operator identification; the identification of the antithetical feature points according to the computing power restrictions and precision requirement of the SIFT operator, accelerate the sound characteristic, or through a HOG a pixel with its surrounding field of a sufficient number of pixels in the difference between the larger mode, determining said pixel may be the corner of the FAST operator; the boundary line of using Canny algorithm.
In the above-mentioned scheme, by using the Harris operator, using his judgment and the edge of the angular measure or the quality of the response, the response function value can be used for the selection of the size of the isolated angular pixel or refinement of the edge pixel, thus the accurate and rapid identification corner point; SIFT (Scale Invariant Feature Transform) operator is the field of computer vision is very well-known characteristics of the operator, it can be used for pattern recognition and Image matching, is a based on the scale space, to Image scaling, rotating even affine transformation does not keep the denaturation of the Image local feature description of the operator that the scale invariant feature transform; acceleration sound i.e. SURF (Speeded Up Robust Features) is a sound Image recognition and described algorithm, for computer vision tasks, such as article identification and 3 D reconstruction; HOG (Histogram ofOriented Gradient) that is the direction gradient histogram is a computer vision and Image processing in object detection characteristic to the descriptors, HOG characteristic is through calculation and statistical Image local area of gradient direction histogram to construct a feature; FAST operator through a certain pixel with its surrounding field of a sufficient number of pixels in the difference between the larger mode, determining said pixel may be angular, in order to solve the problem of detecting the real-time nature of the system; [...] identification according to the computing power of the feature points of the different calculation accuracy and for different needs, to select the algorithm in order to reach the best; Canny edge detection algorithm is a standard algorithm, it is used for boundary identification is limited as much as possible in the Image identifies the actual edge of the, undetected real edge of the probability and the probability of detection of the non-edge are as small as possible.
In one preferred embodiment, step 5 in the orbit fitting information acquisition process are as follows:
Using the Harris operator of said corner point identification.
The gradient descent, Gaussian process or EM algorithm to the curve fitting.
In the above-mentioned scheme, by using the Harris operator, using his judgment and the edge of the angular measure or the quality of the response, the response function value can be used for the selection of the size of the isolated angular pixel or refinement of the edge pixel, thus the accurate and rapid identification corner point, with the step 1 in the similar identification algorithm of the corner spot; re-adopting the step 2 in the curve fitting algorithm with the track curve fitting in, in order to realize low power consumption and high response speed demand.
In one preferred embodiment, step 2 in fitting the edge boundaries of the [...]; the curved surface and the curved surface on the curve fitting using the gradient descent, Gaussian process or EM algorithm.
In the above-mentioned scheme in, through the use of the Hough transform to the data points on the boundary line fitting convenient to reject the interference point, and will be distributed in different boundary line in the vicinity of the separating point, thus reducing the fitting error; gradient descent algorithm is a gradient descent method to the iterative solution step by step, to minimize the loss of function and the model parameter or the need for solving the maximum value of the loss function, on the need to gradient rise method to iterative; Gaussian process is probability and mathematical statistics in a random process, is a series of subordinate to the normal distribution of the random variable in a index collecting inner combined; EM algorithm is an iterative algorithm, on the basis of statistics in is used for searching, can not be observed depending on the probability of hidden variables in the model, the maximum likelihood estimate of parameter; of the above algorithm are not highly accurate algorithm, but meet the fitting precision the premise, can reduce the power consumption and improve the response speed.
In one preferred embodiment, step 6 in the metric can to said curved surface after slicing the use of L2 distance, or based on the probability distribution of the dispersion of the KL divergence or JS.
In the above-mentioned scheme in, by determining the distance metric is based on probability distribution or divergence, it is easy to raise the scene re-identification of the optimized result with the ultrasonic sensor sensing of degree of the scene.
In one preferred embodiment, step 6 in the optimizing algorithm is the gradient drop law.
In the above-mentioned scheme, gradient descent algorithm is a gradient descent method to the iterative solution step by step, to minimize the loss of function and the model parameter or the need for solving the maximum value of the loss function, on the need to increase the iteration method to gradient; a gradient descent method is in order to optimize the loss function, the most commonly used method for optimizing algorithm.
1. The use of high-precision laser radar or other way prebuilt scene map.
2. Extracted from the map information in the scene or object of the curved surface, boundary line, corner point, such as semantic feature points can be used for the map matching contrast characteristic. The identification of the corner spot Harris operator can be used. The semantic feature point marks can be according to the computing power limitation and precision requirement is selected and used SIFT operator, SURF operator, FAST operator HOG operator or the like. The border line of the Canny edge detection identification can be used algorithm.
3. Paragraph 2 in step boundary line and curved surface and fitting. Boundary line fitting can use [...]. Curves and surfaces of the fitting can be used based on the gradient descent, Gaussian process or EM algorithm of the optimization algorithm.
4. The use of efficient database storage section 2 step extracting the feature points and 3 after [...] curved surface information.
5. In the absence of people and vehicles in the form of recording at any time in the process space of the ultrasonic sensor reading of the change of the position of the obstacle information.
6. The process according to the unmanned vehicle in the form of attitude and attitude relative to the body of the ultrasonic sensor to the article 5 step in sensor position information of the coordinate transformation.
7. Article 6 step after the transformation in the position information of the ultrasonic sensor changes over time trajectory similar article 2 in step corner point identification and 3 in step curve fitting.
8. The article 7 step identified in the corner spot and after fitting the curve with the article 4 in step feature point and curved surface matching the information of the comparison of the characteristic points, curves and surfaces matching comparison can be abstract mathematical optimization curve and flow pattern of the difference metric, its metric can be to curved surface after slicing the use of L2 distance, or based on the probability distribution of the dispersion of the KL divergence or JS. The optimization algorithm can be adopt gradient drop law. After the convergence of the optimization algorithm results mean that with the current vehicle consists of an ultrasonic sensor to sense the environment of the scene most matching of the local area.
Although the embodiments of the invention have been disclosed above, but not limited to the specification and embodiments set out in the application, it is fully can be suitable for various suitable for the field of this invention, familiar in the case of the field personnel, can be easily achieved in addition changes, therefore without departing from the claims and the equivalent of the scope of the general concept of defined, the invention is not limited to the specific details shown and described here of the legend.
Scene re-identification algorithm based on ultrasonic sensor
400-4929-0909
contact@catarc.com.cn
