Lidar-based three-dimensional target recognition method

Abstract

The application relates to a kind of three-dimensional target identification methods based on laser radar, first is to construct point cloud model library, and the model point cloud is extracted using ISS algorithm key point;Then global target identification is carried out;On this basis, whether chi-square distance is less than the set threshold value is judged, if initial identification fails, that is, chi-square distance does not have the target meeting the requirement, turn into using local feature descriptor to carry out target identification, the key point of scene point cloud is extracted and PDSH feature is calculated, and local feature matching is carried out using Euclidean distance, finally constructs descriptor, carries out error matching pair elimination to corresponding feature descriptor and votes and identifies the final target.The application realizes local feature extraction of point cloud by the proposed PDSH local feature descriptor, and realizes three-dimensional target identification by combining the proposed PDSH local feature descriptor with GASD global feature descriptor, which guarantees the correctness of identification and speeds up the identification speed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer vision and object recognition, and relates to a three-dimensional object recognition method based on a laser radar. BACKGROUND

[0002] Current object recognition technology mainly includes two-dimensional image-based and three-dimensional image-based, among which the two-dimensional image-based object recognition technology has been quite mature, but it cannot truly and completely express the three-dimensional characteristics of the target in the real scene. In contrast, three-dimensional images can express the actual size and morphological structure of the target in the real world, so the application of three-dimensional images to the field of object recognition makes the accuracy of object recognition greatly improved, and has great advantages.

[0003] In three-dimensional images, point cloud, as a most original three-dimensional data representation form, is a set composed of a plurality of discrete, unordered and non-topological three-dimensional points. In recent years, with the rapid development of hardware and software technology, the emergence of low-cost sensors such as Microsoft Kinect, Google Tango tablet and Intel Real Sense makes the acquisition of point cloud data as convenient as image, further promoting the development of three-dimensional object recognition technology (3D Object Recognition Technology). Current mainstream point cloud data sensors are divided into two categories: active and passive. Laser radar is a sensor that actively detects with laser beams, which can obtain three-dimensional information of the target, including position, attitude and even shape characteristics, and has the advantages of good directivity, high ranging accuracy and all-day working.

[0004] Three-dimensional object recognition technology is not only widely used in the field of automatic driving, but also has important applications in intelligent robots, face recognition, cultural relic restoration, precision guidance and biological recognition systems. Especially in the field of automatic driving, with the help of vehicle-mounted laser radar, a large amount of three-dimensional point cloud data can be obtained to realize accurate recognition of road targets. For example, BMW, Ford and Huawei have all joined the ranks of automatic driving. In the field of intelligent robots, if three-dimensional recognition technology is used, the robot can better guide the intelligent work of the robot using laser radar. In the field of precision guidance and automatic target recognition, three-dimensional images obtained by laser radar can realize target detection and recognition under the shielding of small forests and camouflage nets.

[0005] According to the different scales of the feature descriptors used, mainstream point cloud recognition methods can be divided into two categories: point cloud recognition methods based on global features and point cloud recognition methods based on local features. The global feature descriptor of a point cloud is a descriptor generated for the target as a whole, which has strong descriptive ability for the overall information of the object and is considered to be a representation of the object's geometry. The representative global feature descriptor GASD first constructs a local coordinate system based on principal component analysis, then aligns the point cloud to the local coordinate system, and then establishes a regular grid in space and counts the number of 3D points in each grid. GASD also uses three-line interpolation to eliminate the influence of adjacent grids, and can combine color information to enhance its description ability. Through verification, the GASD descriptor can achieve efficient target estimation and accurate pose estimation, but it is sensitive to local occlusion because it describes the target as a whole. The local feature descriptor of a point cloud is a process of converting the shape information of the key point neighborhood into a high-dimensional vector, and each point is associated with a descriptor. The TOLDI local feature descriptor, after constructing a local reference coordinate system, first transforms the local surface to the reference coordinate system, then projects the local surface to the coordinate plane to form three projection height maps, and finally concatenates the three sub-feature vectors to form the final feature descriptor. Its description is satisfactory, but its dimension is as high as 1200, resulting in low matching efficiency. Therefore, the current use of global or local single descriptors for three-dimensional target recognition has the disadvantages of large amount of calculation, low time efficiency, and poor robustness to clutter and occlusion. SUMMARY

[0006] The purpose of the present application is to propose a laser radar-based three-dimensional target recognition method to solve the problems of large amount of calculation, low time efficiency, and poor robustness to clutter and occlusion existing in the prior art.

[0007] In order to achieve the purpose of the present application, the technical solution adopted by the present application is a laser radar-based three-dimensional target recognition method, which first constructs a point cloud model library, and extracts key points from the model point cloud using the ISS algorithm, calculates the GASD global feature of the model point cloud and the PDSH local feature based on point distribution statistics descriptor, and stores them in the library. Then, global target recognition is performed, the GASD global feature of the scene point cloud is calculated, and the chi-square distance is used for global feature matching with the model library model to realize fast initial recognition of the object. On this basis, it is judged whether the chi-square distance is less than a set threshold value. If the initial recognition fails, that is, there is no target that meets the requirements of the chi-square distance, the local feature descriptor is used for target recognition, the key points of the scene point cloud are extracted and the PDSH feature is calculated, and the Euclidean distance is used for local feature matching. Finally, the descriptor is constructed, the corresponding feature descriptor is eliminated, and the final target is recognized through voting.

[0008] Furthermore, the specific steps for describing local features include:

[0009] 1) Construction of a simplified local coordinate system: The local coordinate system (LRF) is represented as:

[0010] L p ={x (p) x (p) ×z (p) , z (p)}

[0011] Given a keypoint p in a point cloud M and a neighborhood subset P = {p1, p2, ..., p...} determined by the support radius R. k}, then weighted C (p) The calculation is as follows:

[0012]

[0013] The direction of the x-axis is determined by summing the x-coordinates of multiple points in the neighborhood, and the direction in which the sum of the coordinates is greater than zero is defined as the positive direction.

[0014]

[0015] in, Similarly, the direction of the z-axis is determined, and the y-axis and its direction are determined by x×z.

[0016] 2) Surface P / is oriented at a set of angles γ i Simultaneously rotate around the three coordinate axes of the LRF, where γ i The rotation matrix corresponding to i×γ (where γ is the initial rotation angle) is expressed as:

[0017]

[0018] Furthermore, the descriptor construction steps are as follows:

[0019] 1) Generation of sub-eigenvectors of the xy, xz, yz planes

[0020] After rotation, a set of local surfaces is obtained. Will Projected onto the three coordinate planes (xy, xz, yz) of the LRF, the three-dimensional surface is further transformed into a two-dimensional plane;

[0021] The process of generating sub-feature vectors in the xOy plane is as follows: For the projected image in the xOy plane... Project the feature point p to the point p / xOy (γ i Establish a rectangular coordinate system xp with the origin as the origin. / xOy (γ i Then, taking the x-axis as the starting point, and using an angle θ = 360 / N, [the following is a partial equation / parameter]. The plane is divided into N equal-angle regions. Each region contains several projection points, and the farthest projection point on the dividing line between any two adjacent regions is defined as the edge point.

[0022] First, the edge points are used. With feature projection point p / xOy (γ i The distance between the points is used to describe the features. Then, the projection points in the segmented region are used. The number of projection points in each segmented region is counted. Finally, the sub-features of the edge points and the projection points in the segmented region are concatenated to obtain the sub-feature vector of the xOy plane.

[0023] The statistical information is normalized using the support radius and the maximum number of projection points, and expressed as follows:

[0024] f xOy = [vec(d), vec(n)]

[0025]

[0026]

[0027] in, p / xOy (γ i ) and edge points The distance between them, Qty xoy (N i ) is the Nth i The number of projection points in each partitioned region, max(Qty) xOy (N k )) is the maximum number of projection points in all partitioned regions;

[0028] The process of generating sub-eigenvectors in the xOz and yOz planes is the same as that of generating sub-eigenvectors in the xOy plane.

[0029] 2) Concatenate the sub-feature vectors of the xy, xz, and yz planes to obtain the final feature vector, which is the local feature descriptor, represented as:

[0030] f PDSH =[f xOy f xOz f yOz ].

[0031] Furthermore, the key point extraction steps are as follows:

[0032] Suppose there are n points in a point cloud P, and any point p... i Coordinates are (x i y i , z i ), i = 0, 1, ..., n-1. The specific steps are as follows:

[0033] 1) For each point p in the point cloud P i Establish a local coordinate system and set a search radius r for all points;

[0034] 2) Search for each point p in the point cloud i Find all points within the radius r neighborhood and calculate their weights, i.e.

[0035]

[0036] weight —This represents each point P within the neighborhood of radius r. i The weights; where ||P i -P j || represents a point within radius r.

[0037] 3) Calculate each point p i The covariance matrix, i.e.

[0038]

[0039] covariance matrix — Represents the neighborhood of radius r, P i The covariance matrix is ​​T, where T indicates that the matrix is ​​the transpose.

[0040] 4) Calculate the eigenvalues ​​of each covariance matrix. And arrange them in descending order;

[0041] 5) Set thresholds ε1 and ε2, where the thresholds do not exceed 1, satisfying the condition. and The points in the two equations are the key points.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] 1. The PDSH local feature descriptor is used to extract the local feature of the point cloud by counting the distribution information of the points after the rotation projection of the point cloud. Figure 1 The PRC curve of the PDSH descriptor is very close to the longitudinal axis of the coordinate axis, which shows that it has excellent description.

[0044] 2. The PDSH local feature descriptor is used to extract the local feature of the point cloud by counting the distribution information of the points after the rotation projection of the point cloud. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The result of the descriptive test on the point cloud data set is shown in the figure;

[0046] Figure 2 The overall flow chart of the point cloud target recognition is shown in the figure;

[0047] Figure 3 The schematic diagram of the descriptor construction process is shown in the figure;

[0048] Figure 4 The result of the target recognition based on the global feature descriptor is shown in the figure;

[0049] Figure 5 The result of the local feature matching is shown in the figure. DETAILED DESCRIPTION

[0050] The application will be described in detail below in combination with the drawings and specific embodiments.

[0051] The application is aimed at the three-dimensional target recognition technology, and the process of the global initial recognition of the three-dimensional target by using the GASD global feature descriptor is added before the target recognition based on the proposed PDSH local feature descriptor, so as to speed up the recognition efficiency.

[0052] The design principle is that firstly, a point cloud model library is constructed, and key points of the model point cloud are extracted by using an ISS algorithm, and a GASD global feature of the model point cloud and a PDSH (Point Distribution Signature Histogram) local feature based on point distribution are calculated and stored in the library; then global target recognition is performed, the GASD global feature of the scene point cloud is calculated, and chi-square distance is used for global feature matching with the model library model, so that fast initial recognition of the object is realized; on this basis, it is judged whether the chi-square distance is less than a set threshold value, if the initial recognition fails, that is, there is no target meeting the requirement of the chi-square distance, the local feature descriptor is used for target recognition, the key points of the scene point cloud are extracted and the PDSH feature is calculated, and the Euclidean distance is used for local feature matching; finally, the descriptor is constructed, the corresponding feature descriptor is matched and eliminated, and the final target is recognized by voting.

[0053] Embodiment: The application provides a three-dimensional target recognition method based on a laser radar, referring to Figure 2 , and the specific steps are as follows:

[0054] Step one, key point extraction: assuming that there are n points in the point cloud P, any point p i has coordinates (x i , y i , z i ), i=0, 1, …, n-1. The specific steps are as follows:

[0055] 1) a local coordinate system is established for each point p i in the point cloud P, and a search radius r is set for all points;

[0056] 2) all points in the radius r neighborhood of each point p i in the point cloud are searched, and the weight value is calculated, that is,

[0057]

[0058] The weight value represents the weight value of each point P i in the radius r field; wherein ||P i -P j || represents the points in the radius r.

[0059] 3) the covariance matrix of each point p i is calculated, that is,

[0060]

[0061] The covariance matrix represents the covariance matrix of P iCovariance matrix of the point cloud, T represents the transpose matrix.

[0062] 4) Calculate the eigenvalue of each covariance matrix and arrange them in descending order;

[0063] 5) Set threshold values ε1 and ε2, the threshold value does not exceed 1, satisfying and The points of the two formulas are the key points.

[0064] Step two, global target recognition

[0065] The global feature-based point cloud target recognition is to use the global feature descriptor to describe the global features of the scene and model point clouds respectively, and finally obtain two statistical histograms, and further convert the global feature matching into comparing the similarity of the two histograms. For the comparison of histogram similarity, the chi-square distance is usually used to measure, because the histogram point cloud descriptor is usually a multidimensional vector and a probability distribution, and the chi-square distance can be used as the distance measure between two independent variables and the distance measure between two probability distributions of the same dimension. Finally, according to the calculated chi-square distance value, it is compared whether it is less than the set threshold value, and the model point cloud with the smallest distance and far less than the set threshold value is taken as the final recognized result, as shown in Figure 3 .

[0066] Step three, local feature description:

[0067] The local feature description is the most critical step of the point cloud target recognition system, which is to calculate the geometric features of each key point neighborhood for description. The specific steps are as follows:

[0068] 1) Construction of a simplified local coordinate system: the local coordinate system LRF is represented as:

[0069] L p ={x (p) ,x (p) ×z (p) ,z (p)}

[0070] Given a key point p in the point cloud M and a neighborhood subset P = {p1, p2,..., p k} determined according to the support radius R, the calculation of the weighted C (p) is as follows:

[0071]

[0072] To eliminate the ambiguity of LRF direction, the sum of x coordinates of the points in the neighborhood is used to determine the direction of x axis, and the direction of which the sum of coordinates is greater than zero is defined as positive direction.

[0073]

[0074] wherein, The direction of z axis is determined in the same way, and the direction of y axis is determined by x x z.

[0075] 2) The curved surface P / is rotated by a set of angles γ i around the three coordinate axes of LRF, wherein γ i = ix γ (γ is the initial rotation angle) corresponds to the rotation matrix:

[0076]

[0077] Step four, the construction of the descriptor is as follows:

[0078] 1) Generation of sub-feature vectors on x-y, x-z, y-z planes

[0079] After rotation, a set of local curved surfaces are obtained are projected onto the three coordinate planes (x-y, x-z, y-z) of LRF, and further converted into two-dimensional planes.

[0080] Take the process of generating sub-feature vectors on xOy plane as an example, as shown in Figure 4 , for the projected image on xOy plane, the projection point p / xOy (γ i ) of feature point p is taken as the origin to establish a rectangular coordinate system xp / xOy (γ i )y. Then, the x axis is taken as the starting point, and the xOy plane is divided into N equal angle regions by an angle θ = 360 / N, and there are several projection points in each region, and the farthest projection point on the dividing line between each two adjacent regions is defined as the edge point

[0081] Firstly, the edge point is used to determine the direction of the feature projection point p / xOy (γ i ​​The distance between points is used for feature description, followed by the projection points within the segmented regions. The number of projection points within each segmented region is counted. Finally, the sub-features of the edge points and the projection points within the segmented regions are concatenated to obtain the sub-feature vector of the xOy plane. To achieve strong robustness to resolution changes, the statistical information is normalized using the support radius and the maximum number of projection points, expressed as:

[0082] f xOy = [vec(d), vec(n)]

[0083]

[0084]

[0085] in, p / xOy (γ i ) and edge points The distance between them, Qty xOy (N i ) is the Nth i The number of projection points in each partitioned region, max(Qty) xOy (N k )) is the maximum number of projection points in all partitioned regions.

[0086] The process of generating sub-eigenvectors in the xOz and yOz planes is the same as that of generating sub-eigenvectors in the xOy plane.

[0087] 2) Concatenate the sub-feature vectors of the xOy, xOz, and yOz planes to obtain the final feature vector, which is the local feature descriptor, represented as:

[0088] f PDSH =[f xOy f xOz f yOz ]

[0089] At this point, the new descriptor construction is complete, and the overall recognition flowchart is as follows: Figure 2 As shown, the recognition effect is as follows: Figure 5 As shown.

[0090] The new descriptors obtained in the above embodiments are compared with existing descriptors through target identification to obtain, as follows: Figure 1 The line graph shows that the Precision-Recall Curve (PRC) is used to evaluate the discriminative power of feature descriptors. The horizontal axis, 1-precision, in the PRC curve is defined as the ratio of incorrect matches to the total number of matches.

[0091]

[0092] The recall is defined as the ratio of the number of correctly matched pairs to the corresponding number of matched pairs:

[0093]

[0094] By changing the threshold τ f A plurality of recall rates and their corresponding precisions can be obtained, thereby obtaining a PRC curve. For an ideal feature descriptor, its PRC curve should be concentrated in the upper left corner of the rectangular coordinate system.

[0095] The above is a description of the specific implementation of the present application, rather than a limitation of the present application. Those skilled in the art can make various equivalent technical solutions without departing from the scope of the present application, and therefore all equivalent technical solutions should be included in the patent protection scope of the present application.

Claims

1. A method of laser radar-based three-dimensional target recognition, characterized by: Firstly, a point cloud model library is constructed, and key points are extracted from the model point cloud using the ISS algorithm, the GASD global feature of the model point cloud is calculated, and the PDSH local feature based on point distribution is stored in the library; then global target recognition is performed, the GASD global feature of the scene point cloud is calculated, the chi-square distance is used for global feature matching with the model library model, and fast initial identification of objects is realized; on this basis, it is judged whether the chi-square distance is less than the set threshold value, if the initial identification fails, that is, there is no target meeting the requirements of the chi-square distance, the local feature descriptor is used for target recognition, the key points of the scene point cloud are extracted and the PDSH feature is calculated, and the Euclidean distance is used for local feature matching, finally the descriptor is constructed, the corresponding feature descriptor is matched and eliminated, and the final target is recognized through voting identification; The construction steps of the descriptor are as follows 1) Generation of sub-feature vectors of x-y, x-z and y-z planes A set of local surfaces is obtained after rotation The Projecting to three coordinate planes (x-y, x-z, y-z) of LRF, further converting three-dimensional surface into two-dimensional plane; The process of generating the sub-feature vector of the xOy plane is: for the projection image in the xOy plane The projection point p of the feature point p / xOy (γ i ) as the origin to establish a rectangular coordinate system xp / xOy (γ i ) y , then, taking the x-axis as the starting point, the xOy plane is divided into N equal angle regions at an angle θ = 360 / N, and there are several projection points in each region, and the farthest projection point on the division line between each two adjacent regions is defined as an edge point ​ Firstly, the edge points are described by the distance between each edge point and the feature projection point p / xOy (γ i ) between each edge point and the feature projection point p, then the projection points in the divided region are described by counting the number of the projection points in each divided region, finally, the sub-features of the edge points and the projection points in the divided region are concatenated to obtain the sub-feature vector of the xOy plane.​ The statistical information is normalized by using the support radius and the maximum number of projection points, and is expressed as: f xOy = [vec(d), vec(n) wherein, represents p / xOy (γ i ) is the distance between the edge point and the projection point, Qty xOy (N i ) is the number of projection points in the N i th division region, and ma(Qty xOy (N k )) is the maximum number of projection points in all division regions. The process of generating sub-feature vectors of xOz and yOz planes is the same as that of generating sub-feature vectors of xOy plane; 2) The sub-feature vectors of x-y, x-z and y-z planes are spliced to obtain the final feature vector, and the local feature descriptor is obtained, and is expressed as: f PDSH = [f xOy ,f xOz ,f yOz ].

2. The laser radar-based three-dimensional object recognition method according to claim 1, characterized by: The specific steps of the local feature description include: 1) Construction of a simplified local coordinate system: the local coordinate system LRF is expressed as: L p = {x (p) ,x (p) ×z (p) ,z (p)} Given a key point p in the point cloud M and a neighborhood subset P = {p1, p2,..., p k} determined according to a support radius R, the computation of the weight C (p) is as follows: The direction of the x-axis is determined by using the sum of x coordinates of multiple points in the neighborhood, and the direction with a coordinate sum greater than zero is defined as the positive direction wherein The direction of the z-axis is determined analogously, the y-axis and its direction are determined by x x z. 2) the curved surface P / at a set of angles γ i while rotating around the three coordinate axes of the LRF, where γ i = i x γ, γ is the initial rotation angle, and the corresponding rotation matrix is represented as:

3. The laser radar-based three-dimensional object recognition method according to claim 1 or 2, characterized by, The steps of the key point extraction are as follows: Suppose that there are n points in the point cloud P, any point p i The coordinates of the point are (x i , y i , z i ), i = 0, 1, …, n-1, and the specific steps are as follows: 1) For each point p in point cloud P i A local coordinate system is established and a search radius r is set for all points; 2) Search for each point p in the point cloud i All points within the radius r neighborhood, and compute their weights, i.e. Equation (4) represents the weight value at each point P i in the field of radius r; wherein ||P i -P j || represents a point within a radius r; 3) Compute the covariance matrix of each point p i i.e. Equation (5) represents the covariance matrix of the field P at radius r i T represents that the matrix is transposed. 4) Calculate eigenvalues of each covariance matrix and arrange in descending order; 5) Set threshold values ε1 and ε2, threshold values not more than 1, satisfying and The points of the two formulas are the key points.

Images