Multiscale point cloud classification method and system
The present disclosure discloses a multiscale point cloud classification method. The method includes the following steps: acquiring 3D unordered point cloud data; performing feature extraction and classification on the acquired point cloud data using a pre-trained parallel classification network to obtain an output result, wherein the parallel classification network includes a plurality of basic networks with the same structures; and fusing the output results of the parallel network using a pre-trained deep Q network to obtain a final result of point cloud classification. The present disclosure can improve the accuracy and robustness of point cloud classification.
1. A multiscale point cloud classification method, applied on recognition of an object in three-dimensional (3D) space, comprising:
acquiring 3D unordered point cloud data;
performing feature extraction and classification on the acquired point cloud data using a pre-trained parallel classification network, and obtaining an output result, wherein the parallel classification network comprises an input layer, and an output layer composed of five basic networks with the same structures, wherein the various basic networks are distinguished according to different feature dimensions of point clouds input into the basic networks, and the dimensions of features output by different layers in the same basic network are the same; wherein each basic network includes a feature alignment layer, a feature extraction layer and a fully connected layer in sequence; the feature alignment layer is used for learning point cloud position features; the feature extraction layer is used for extracting global features; the fully connected layer is used for outputting results; and
fusing the output results of the parallel network using a pre-trained deep Q network to obtain a final result of point cloud classification, to determine various objects in the 3D space.
2. The multiscale point cloud classification method according to claim 1 , wherein the obtaining an output result comprises:
performing feature alignment on the point clouds by means of the feature alignment layer according to point cloud data of n×3;
after a convolution operation is performed on the aligned point cloud data of n×3 for three times through the feature extraction layer, outputting features with dimensions of n×128, n×256, n×512, n×1024 and n×2048;
after max-pooling operation, solving a maximum value at the second dimension, and outputting global features with dimensions of 1×128, 1×256, 1×512, 1×1024 and 1×2048; and
performing processing through the fully connected layer, and obtaining scores of k categories with a size of N×k as the output results of the basic networks, N being a batch size.
3. The multiscale point cloud classification method according to claim 1 , wherein the performing, by the feature extraction layer, feature extraction using a feature pyramid network comprises:
extracting features according to an order from low dimensions to high dimensions, performing double upsampling and addition on the extracted features in order from low dimensions to high dimensions, and outputting features with corresponding feature dimensions.
4. The multiscale point cloud classification method according to claim 1 , wherein the various basic networks that constitute the parallel network are trained using a loss function, and the loss function is expressed by the following formula:
l
(
θ
)
=
-
∑
i
=
1
m
log
(
P
G
(
x
i
;
θ
)
)
(
1
)
in formula (1), xi represents a sample; m represents the number of training samples; and PG (xi; θ) represents a data distribution model of each sample decided by parameter θ.
5. The multiscale point cloud classification method according to claim 1 , wherein a dataset for training the deep Q network is an object category with the best classification result achieved by manually labeling each basic model.
6. A computing device, comprising a processor and a storage medium, wherein the storage medium is configured for storing instructions; and
the processor is configured for performing, according to the instructions, the steps of the method according to claim 1 .