Target point cloud efficient identification method based on dynamic switching

Abstract

The invention discloses a target point cloud efficient identification method based on dynamic switching, relates to the field of computer graphics and three-dimensional point cloud target identification, and solves the problem that the existing target point identification is single in method and insufficient in stability. According to the method, through combination of block identification, accuracy judgment and secondary identification, primary identification is carried out on a target point cloud formed by a laser radar through a DGCNN deep learning part segmentation method, the identification efficiency is optimized by setting priorities, and according to different target identification accuracy rates, an identification method based on feature matching is dynamically switched, so that the identification efficiency is improved. According to the point cloud target recognition method, the recognition accuracy of the point cloud target is improved, and the false alarm rate is reduced.

Description

Technical Field

[0001] The present invention relates to the fields of computer graphics and three-dimensional point cloud target recognition, and in particular to an efficient target point cloud recognition method based on dynamic switching. Background Art

[0002] With the rapid development of science and technology, unmanned systems have emerged in fields such as autonomous driving, intelligent navigation, and low-altitude economy. Three-dimensional imaging lidar (LiDAR) is a sensor that uses laser beams for active detection. It can acquire three-dimensional point cloud images of targets and possesses high-precision imaging detection capabilities around the clock. The use of LiDAR enables three-dimensional imaging of targets and has broad application prospects. However, due to factors such as illumination variations, deformation, occlusion interference, and multiple model variants in complex scenes, a single recognition method often fails to meet the dual requirements of generalization and accuracy in practical engineering. The deep learning-based DGCNN algorithm offers high recognition accuracy and speed. However, due to the inherent generalization performance limitations of deep learning models, this can lead to a significant decrease in recognition accuracy. Therefore, a composite research framework combining deep learning and feature matching is adopted. By integrating the advantages of these two methods, a more robust recognition system is constructed, breaking through the traditional single-modal recognition paradigm and addressing the challenges of point cloud target recognition and positioning in complex situations. Summary of the Invention

[0003] The present invention provides an efficient target point cloud recognition method based on dynamic switching to solve the problem of insufficient stability of the existing single target point recognition method.

[0004] An efficient target point cloud recognition method based on dynamic switching is implemented by the following steps:

[0005] Step 1: Use the DGCNN model to segment the target point cloud generated by the lidar;

[0006] Step 2: Prioritize the segmented parts and calculate the segmentation accuracy of the highest priority;

[0007] Step 3: Determine whether the accuracy reaches the set threshold. If so, calculate the target point; otherwise, identify the target through a feature matching-based method and calculate the final target point.

[0008] Beneficial effects of the present invention:

[0009] 1. The present invention prioritizes target areas, reduces computational redundancy, and improves algorithm efficiency.

[0010] 2. The present invention uses the deviation angles between the connection vector and each coordinate axis of the LRF and the deviation angles between the connection vector and the point cloud normal vector to effectively characterize the target point cloud. Finally, the target with recognition accuracy lower than the threshold is re-identified through the feature matching method of the nearest neighbor distance ratio to obtain the final target point coordinates.

[0011] 3. The algorithm proposed in this paper is experimentally verified on the Shapenetpart dataset. The algorithm proposed in this paper can accurately identify the parts of automobile targets. Compared with pointnet and pointnet++, the false alarm rates are reduced by 0.86% and 0.73% respectively, and the false alarm rate is 0%. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 This is a flow chart of the method for efficiently identifying target point clouds based on dynamic switching according to the present invention;

[0013] Figure 2 This is the DGCNN model architecture diagram;

[0014] Figure 3 Schematic diagram of the deviation angle between the connection vector and LRF;

[0015] Figure 4 Schematic diagram of the deviation angle between the connection vector and the point cloud normal vector;

[0016] Figure 5 This is the target point recognition result map with the vehicle head as the highest priority;

[0017] Figure 6 This is the target point recognition result diagram with the vehicle body as the highest priority. DETAILED DESCRIPTION

[0018] Specific implementation method 1. Combination Figures 1 to 4 This embodiment describes an efficient target point cloud recognition method based on dynamic switching, such as Figure 1 As shown, the specific steps include:

[0019] Step S1: Use the DGCNN model to segment the target point cloud formed by the laser radar to obtain the segmentation points and classification results of each part; the specific process is as follows:

[0020] Step S11: Figure 2 As shown in the figure, after layer-by-layer graph convolution and pooling, a global description is obtained, and then the classification result c of the target point cloud is obtained through a multi-layer perceptron;

[0021] Step S12: By stacking graph convolution operations and residual connection operations, the global features are finally obtained through pooling operations, and then spliced ​​with the previously extracted features and input into the multi-layer perceptron to obtain a segmentation result of n×p shape.

[0022] Step S2: Figure 1 As shown, according to the segmentation points of each part obtained in step S1, the priority ranking of the segmented parts is set to determine whether it is priority part 1. If so, the segmentation accuracy is calculated; otherwise, it is determined whether it is priority part 2. If so, the segmentation accuracy of priority part 2 is calculated; otherwise, it is determined whether it is priority part 3. If so, the segmentation accuracy of priority part 3 is calculated. Similarly, subsequent parts are determined in turn.

[0023] Step S21, set the part set C = {c1, c2, ..., c N ,…,c M}, define the priority function:

[0024]

[0025] In the formula, P(c N ) indicates the part c N priority.

[0026] Identify the highest priority areas:

[0027]

[0028] Where c * The highest priority area.

[0029] Step S22: Calculate the segmentation accuracy according to formula (8), which is:

[0030]

[0031] Where M is the number of parts; y n is the true part category of the nth point cloud in part N; is the predicted part category of the nth point cloud in part N; φ is the indicator function, which is 1 when the prediction is correct and 0 otherwise.

[0032] Step S3: Determine whether the accuracy of the highest priority part in the DGCNN model segmentation result reaches a threshold. If so, execute step S4; otherwise, execute step S5;

[0033] In this embodiment, the determination condition for determining whether the accuracy of the highest priority part reaches the threshold θ is set as:

[0034] A(c * )>θ (9)

[0035] Among them, A(c * ) is the highest priority part c of the DGCNN model * segmentation accuracy.

[0036] Step S4: Calculate the target point coordinates:

[0037]

[0038] Where A∈[0,1] is the current accuracy; θ∈[0,1] is the threshold; X is the data point set; P c is the center point; d(x,P c ) is the distance from point x to the center point P c F(X) is the target point based on the feature matching method.

[0039] Step S5: segment the parts and calculate the target points using a feature matching method;

[0040] Step S51: sampling feature points of the target point cloud and constructing a local reference coordinate system;

[0041] First, the PCA method is used to sample feature points of the target point cloud and construct a local spherical neighborhood space;

[0042] Calculate the normal vector of each point in the point cloud, let P i For any point in the point cloud P, construct the covariance matrix of the point and its semi-positive definite matrix Cov(P i );

[0043] By calculating Cov(P i )'s eigenvalues ​​and eigenvectors to obtain P i Normal vector The calculated As P i The main direction, and assign a weight parameter to it

[0044]

[0045] in, It's point P i The curvature of Find all distances P i Less than points, which are recorded as a set for will be collected Middle point p s The normal vector is projected onto a plane superior( is perpendicular to P i A plane connected to the origin), the angle of projection for:

[0046]

[0047] in is the angle before projection;

[0048] Define point P i Interest value I(P i ):

[0049]

[0050] I(P i )=I1(P i )·I2(P i ) (16)

[0051] Where, I1(P i ) and I2(P i ) is point P i Two values ​​of interest, f(P i ) is point P i Function of interest, select all I1(P i ) and I2(P i ) points greater than 0.8 are taken as feature points smp, and the local spherical neighborhood space is constructed;

[0052] Secondly, calculate the local spherical neighborhood space radius of the sampled point cloud and the volume of the local spherical neighborhood

[0053]

[0054] Where, d x , d y , d z is the maximum value of the xyz axis of the local spherical neighborhood, N smp is the number of feature points obtained after sampling, S xyz is the area of ​​the local minimum bounding box.

[0055] Then, by calculating the feature point P di (x Pdi ,y Pdi ,z Pdi )The weight of the nearby point cloud is obtained to obtain the divergence matrix L of the point cloud:

[0056]

[0057] in is a point in the local spherical neighborhood space.

[0058] Finally, perform eigenvalue decomposition on L to obtain three eigenvalues and set up The x-axis direction of the local reference frame (LRF) is:

[0059]

[0060] Where x + is the eigenvector corresponding to the L eigenvalue, x - is x + The reverse eigenvector of Is consistent with x + Elements, Is consistent with x - , j is or Elements in .

[0061] if The ratio of elements in If there are more, the x-axis of LRF is selected as x + direction; otherwise, the direction of the x-axis is x - Select the other axes in the same way. The z-axis is calculated in the same way, and the direction of the y-axis is obtained from x×z. The constructed local reference coordinate system LRF is:

[0062]

[0063] Where A x It determines the direction of the x-axis or A z Similarly; and They are and The number of elements in ; and They are and A z - The number of elements in ; as well as They are the x-axis, y-axis and z-axis of LRF respectively.

[0064] Step S52: Calculate the connection vector between the local spherical neighborhood center of gravity and the origin, and characterize the point cloud features by the deviation angle between the connection vector and the coordinate axis of the local reference coordinate system and the deviation angle of the point cloud normal vector;

[0065] Convert the coordinates of the midpoint of the local sphere neighborhood of the feature point to the corresponding LRF. For example, in and Take any point P on the plane LRFx , The distance D between this plane LRFxIt can be expressed as:

[0066]

[0067] Among them D LRFx for point The x-axis coordinates, y-axis and z-axis coordinates in the LRF are calculated in the same way. Repeat this step for all points in the point cloud so that the coordinates of all points in the local spherical neighborhood of each feature point are transformed into their corresponding LRF.

[0068] The center of gravity Connected to the feature point, the connected vector for:

[0069]

[0070] Where, is a point contained in the local spherical neighborhood of the feature point, Center of gravity The three-dimensional coordinates of are the 3D coordinates of the containing point within the local spherical neighborhood.

[0071] like Figure 3 As shown, calculate the connection vector Deviation angle from LRF:

[0072]

[0073] in and Connection vectors The cosine of the angles of deviation from the LRF's x-axis, y-axis, and z-axis, and The connection vectors Deviation angles from the x-axis, y-axis, and z-axis of the LRF.

[0074] Then, this step is repeated for each feature point of the point cloud, and the deviation angle between each feature point and the LRF is calculated.

[0075] like Figure 4 As shown, the feature points are calculated using the cosine formula All point normal vectors and connection vectors in the local spherical neighborhood of The deviation angle set

[0076]

[0077] Where, for point The normal vector of the point in the local spherical neighborhood of .

[0078] Step S53: Identify the target point by the distance ratio between the nearest neighbor point and the double deviation angle, and calculate the point cloud part with the highest priority;

[0079] Compute the feature histogram of the double deviation angle:

[0080]

[0081] In the formula, Feature xyz It is the deviation angle histogram feature of the feature point and LRF, Feature NV is a feature point All point normal vectors and connection vectors in the local spherical neighborhood of The deviation angle histogram feature of Feature is Feature xyz and Feature NV The merged features are used as the histogram features of the target;

[0082] The histogram feature Feature of the target is matched and compared with the histogram feature in the model library. If the match is successful, the target point is segmented and step S54 is executed. The feature matching is performed using the Euclidean distance method based on the nearest neighbor distance ratio, which is calculated as follows:

[0083]

[0084] In the formula, Feature d is the histogram feature of the target to be identified, Feature m is the histogram feature of the model library.

[0085] Step S54: Calculate the point closest to the center point using the nearest neighbor search method and use it as the target point.

[0086] Calculate the center coordinates of the matching result and search for the nearest neighbor point, taking the nearest neighbor point as the target point:

[0087]

[0088] Where x * is the target point, S is the matching point set, p is the matching point, P c It is the center point.

[0089] Specific implementation method 2: Figure 5 and Figure 6 This embodiment is described as an experimental example of the efficient target point cloud recognition method based on dynamic switching described in the first embodiment.

[0090] To verify the effectiveness of the efficient target point cloud recognition method described in the present invention, this embodiment was experimentally verified on the Shapenetpart dataset, and target points were recognized for the car category in the Shapenetpart dataset. The false alarm rate (FPR) was used to evaluate the performance of the tracking algorithm. The FPR refers to the Euclidean distance between the center position of the true target frame and the center position of the predicted target frame during the recognition process. The smaller the Euclidean distance, the higher the tracking accuracy. The FPR calculation formula is as follows:

[0091]

[0092] Where FPR is the false alarm rate, FP is the number of false positives, and TN is the number of true negatives.

[0093] Shapenetpart dataset experiments:

[0094] The target is the Shapenetpart dataset tracking the "car" category, which is divided into four parts: the front, roof, body, and tire. Set the front position as the highest priority, identify the front target point and mark it with a sphere. The target point recognition is as follows: Figure 5 shown.

[0095] from Figure 5 As can be seen in the figure, in the four different types of vehicle point clouds, the front of the vehicle exhibits a clear and coherent distribution in the grayscale image, and in most samples, the front region is accurately located as the target point area. This demonstrates that under the current recognition strategy, the vehicle front, the highest-priority structure, can be effectively distinguished and recognized. The distribution of target points at the vehicle front remains concentrated and coherent, with no significant false or missed detections, demonstrating strong stability and robustness.

[0096] Figure 6 This is the target point identification result diagram with the roof set as the highest priority using the method of this embodiment. It can be seen in the figure that the roof area is accurately identified as a medium grayscale area, covering the top of the vehicle, and the shape presents a continuous block distribution, which basically matches the roof outline of the real geometric structure. There is no obvious break or separation in this area, indicating that after setting the roof as the priority target, the model can focus on the plane area above the vehicle and form a stable response. Compared with setting the front of the vehicle as the priority target before, this time after setting the roof as the focus, the method of this embodiment can accurately identify the modified priority target point. As shown in Table 1, Table 1 is the comparison result of the false alarm rate between the method of the present invention and the existing method.

[0097] Table 1

[0098]

[0099] As can be seen from Table 1, the false alarm rate of the method described in this embodiment is reduced by 0.86% compared with pointnet and by 0.73% compared with pointnet++, which proves that the method described in this embodiment can accurately identify the target points.

[0100] The technical features of the above-mentioned embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the technical features in the above-mentioned embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. An efficient target point cloud recognition method based on dynamic switching is characterized by: The method is implemented by the following steps: Step 1: Use the DGCNN model to segment the target point cloud generated by the laser radar to obtain the segmentation points and classification results of each part; Step 2: prioritize the segmented parts and calculate the segmentation accuracy of the highest priority part; Step 3: Determine whether the accuracy reaches the set threshold. If so, calculate the target point coordinates; otherwise, identify the target point through a feature matching method and calculate the final target point.

2. The efficient target point cloud recognition method based on dynamic switching according to claim 1 is characterized in that: The specific process of step one is: Step 1: Perform pooling operation after layer-by-layer graph convolution to obtain global features, and then obtain the classification result c of the parts in the target point cloud through multi-layer perceptron; Step 1 and 2: By stacking graph convolution operations and residual connection operations, the global features are finally obtained through pooling operations, and then spliced ​​with the intermediate layer features and input into the multi-layer perceptron to obtain n×p segmentation points.

3. The efficient target point cloud recognition method based on dynamic switching according to claim 1 is characterized in that: In step 2, the segmentation accuracy of the highest priority part is calculated using the following formula: The segmentation accuracy formula is: Where, M is the number of points; y n is the true part category of the nth point cloud; is the predicted part category of the nth point cloud; φ is the indicator function, which is 1 when the prediction is correct and 0 otherwise.

4. The method for efficiently identifying target point clouds based on dynamic switching according to claim 1, characterized in that: In step 3, the condition for determining whether the accuracy of the highest priority part reaches the set threshold θ is: A(c * )>θ In the formula, A(c * ) is the highest priority part c of the DGCNN model * segmentation accuracy.

5. The method for efficiently identifying target point clouds based on dynamic switching according to claim 1, characterized in that: In step 3, the formula for calculating the target point coordinates is as follows: Where A∈[0,1] is the current accuracy; X is the data point set; P c is the center point; d(x,P c ) is the distance from point x to the center point P c F(X) is the target point calculated based on the feature matching method.

6. The method for efficiently identifying target point clouds based on dynamic switching according to claim 1, characterized in that: In step three, the process of identifying the target point based on the feature matching method is as follows: Step 31: Sample feature points of the target point cloud and construct a local reference coordinate system; Step 32: Calculate the connection vector between the local sphere neighborhood center of gravity and the origin, and use the deviation angle between the connection vector and the local coordinate system coordinate axis and the deviation angle of the point cloud normal vector, that is, the double deviation angle; characterize the point cloud features by the double deviation angle; Step 3. Identify the target point by the distance ratio between the nearest neighbor point and the double deviation angle, and calculate the point cloud part with the highest priority; Step 3 and 4: Use the nearest neighbor search to find the point closest to the center point as the target point.

7. The efficient target point cloud recognition method based on dynamic switching according to claim 6, characterized in that: In step 31, the process of constructing the local reference coordinate system is as follows: First, the PCA method is used to sample feature points of the target point cloud and construct a local spherical neighborhood space; Secondly, calculate the local sphere neighborhood space radius and the volume of the local spherical neighborhood Where, d x , d y , d z is the maximum value of the xyz axis of the local spherical neighborhood, N smp is the number of feature points obtained after sampling, S xyz is the area of ​​the local minimum bounding box; Then, calculate the feature points The weights of nearby point clouds are used to obtain the divergence matrix L of the point cloud: Finally, the eigenvalue decomposition of the scatter matrix L is performed to obtain three eigenvalues and set up The x-axis direction of the local reference coordinate system is: Where, is a point in the local spherical neighborhood space, x + is the eigenvector corresponding to the L eigenvalue, x - is x + The reverse eigenvector of Is consistent with x + Elements, Is consistent with x - , j is or Elements in if The ratio of elements in If the x-axis of the local reference coordinate system is x + direction; otherwise, the direction of the x-axis is x - ; Select other axes in the same way. The z-axis is calculated in the same way. The direction of the y-axis is obtained by x×z; the local reference coordinate system LRF is constructed; Where A x It determines the direction of the x-axis or A z Similarly; and They are and The number of elements in ; and They are and The number of elements in ; as well as They are the x-axis, y-axis and z-axis of LRF respectively.

8. The method for efficiently identifying target point clouds based on dynamic switching according to claim 6, characterized in that: The specific process of step 32 is as follows: Convert the coordinates of all points in the local sphere neighborhood of the feature point to the corresponding LRF, and convert the center of gravity Connected with the feature point, the connected vector is Calculate the connection vector The deviation angle from the LRF is expressed as follows: Where, and The connection vectors The cosine of the angles of deviation from the LRF's x-axis, y-axis, and z-axis, and The connection vectors Deviation angles from the x-axis, y-axis, and z-axis of the LRF; The same calculation process as above is used to calculate the deviation angle between each feature point and LRF; Calculate feature points using the cosine formula All point normal vectors and connection vectors in the local spherical neighborhood of The deviation angle set Where, for point The normal vector of the point in the local spherical neighborhood of .

9. The method for efficiently identifying target point clouds based on dynamic switching according to claim 6, characterized in that: In step 33, the characteristic histogram of the double deviation angle is calculated; it can be expressed as follows: Feature=(Feature,Feature NV ) In the formula, Feature xyz is the deviation angle histogram feature between the feature point and LRF, Feature NV is a feature point All point normal vectors and connection vectors in the local spherical neighborhood of The deviation angle histogram feature of Feature is Feature xyz and Feature NV The combined features are used as the histogram features of the target; The target's histogram feature Feature is matched and compared with the histogram feature in the model library. If the match is successful, the target point is segmented. The feature matching calculation is performed using the Euclidean distance method based on the nearest neighbor distance ratio as follows: In the formula, Feature d is the histogram feature of the target to be identified, Feature m is the histogram feature of the model library.

10. The method for efficiently identifying target point clouds based on dynamic switching according to claim 6, characterized in that: In steps 3 and 4, the center coordinates of the matching results are calculated and the nearest neighbor point is searched, and the nearest neighbor point is used as the target point; it can be expressed as follows: Where p * is the target point, S is the matching point set, p is the matching point, P c It is the center point.

Images