THREE-DIMENSIONAL TARGET DETECTION METHOD AND APPARATUS

1 . A three-dimensional target detection method, comprising:

obtaining an image and point cloud data of a target environment;

obtaining semantic information of the image, wherein the semantic information comprises category information corresponding to pixels in the image; and

determining three-dimensional location information of a target in the target environment based on the point cloud data, the image, and the semantic information of the image.

2 . The method according to claim 1 , wherein the determining three-dimensional location information of a target in the target environment based on the point cloud data, the image, and the semantic information of the image comprises:

projecting the image and the semantic information of the image into the point cloud data to generate semantic point cloud data;

extracting feature information of the semantic point cloud data to generate semantic point cloud feature information; and

determining the three-dimensional location information of the target in the semantic point cloud data based on the semantic point cloud feature information.

3 . The method according to claim 2 , wherein the semantic point cloud feature information is output by using a semantic point cloud feature recognition network, and the three-dimensional location information is output by using a target detection network.

4 . The method according to claim 3 , wherein:

the semantic point cloud feature recognition network comprises a point cloud feature recognition subnetwork and an image feature recognition subnetwork;

the point cloud feature recognition subnetwork is used to extract point cloud feature information of the point cloud data; and

the image feature recognition subnetwork is used to extract image feature information of the image based on the image and the semantic information, and dynamically adjust a network parameter of the point cloud feature recognition subnetwork based on the image feature information.

5 . The method according to claim 4 , wherein:

the point cloud feature recognition subnetwork comprises at least one network layer, and the image feature recognition subnetwork is separately connected to each network layer of the at least one network layer; and

the image feature recognition subnetwork is used to extract the image feature information of the image based on the image and the semantic information, and separately and dynamically adjust a network parameter of each network layer of the at least one network layer based on the image feature information.

6 . The method according to claim 4 , wherein the network parameter comprises at least one of a convolution kernel parameter or an attention mechanism parameter, and the attention mechanism parameter is used to determine information that is in the image feature information and whose correlation with the point cloud data is greater than a correlation threshold as valid information for adjusting the point cloud feature recognition subnetwork.

7 . The method according to claim 5 , further comprising:

separately obtaining output data of each network layer of the at least one network layer; and

determining, based on the output data, adjustment effect data corresponding to the network parameter.

8 . The method according to claim 3 , wherein the semantic point cloud feature recognition network and the target detection network are obtained through training in the following manner:

obtaining a plurality of semantic point cloud training samples, wherein the plurality of semantic point cloud training samples comprise a point cloud data sample, an image sample projected to the point cloud data sample, and the semantic information of the image sample, and the three-dimensional location information of the target is labeled in the plurality of semantic point cloud training samples;

constructing the semantic point cloud feature recognition network and the target detection network, wherein an output end of the semantic point cloud feature recognition network is connected to an input end of the target detection network;

separately inputting the plurality of semantic point cloud training samples to the semantic point cloud feature recognition network, and outputting a prediction result by using the target detection network; and

performing iterative adjustment on network parameters of the semantic point cloud feature recognition network and the target detection network based on a difference between the prediction result and the labeled three-dimensional location information of the target, until iteration meets a preset requirement.

9 . The method according to claim 1 , wherein the obtaining semantic information from the image comprises:

performing panoramic segmentation on the image to generate the semantic information of the image, wherein the semantic information comprises a panoramic segmentation image of the image, and the panoramic segmentation image comprises image regions, obtained through panoramic segmentation, of different objects and category information corresponding to the image regions.

10 . The method according to claim 1 , wherein the image comprises a panoramic image.

11 . A three-dimensional target detection apparatus, comprising:

at least one processor; and

a memory coupled to the at least one processor and storing programming instructions for execution by the at least one processor to perform operations comprising:

obtaining an image and point cloud data of a target environment;

obtaining semantic information of the image, wherein the semantic information comprises category information corresponding to pixels in the image; and

determining three-dimensional location information of a target in the target environment based on the point cloud data, the image, and the semantic information of the image.

12 . The apparatus according to claim 11 , wherein the operations comprise:

projecting the image and the semantic information of the image to the point cloud data to generate semantic point cloud data;

extracting feature information of the semantic point cloud data to generate semantic point cloud feature information; and

determining the three-dimensional location information of the target in the semantic point cloud data based on the semantic point cloud feature information.

13 . The apparatus according to claim 12 , wherein the semantic point cloud feature information is output by using a semantic point cloud feature recognition network, and the three-dimensional location information is output by using a target detection network.

14 . The apparatus according to claim 13 , wherein the semantic point cloud feature recognition network comprises a point cloud feature recognition subnetwork and an image feature recognition subnetwork, and the operations comprise:

extracting point cloud feature information of the point cloud data; and

extracting image feature information of the image based on the image and the semantic information, and dynamically adjust a network parameter of the point cloud feature recognition subnetwork based on the image feature information.

15 . The apparatus according to claim 14 , wherein the point cloud feature recognition subnetwork comprises at least one network layer, the image feature recognition subnetwork is separately connected to each network layer of the at least one network layer, and the operations comprise:

extracting the image feature information of the image based on the image and the semantic information, and separately and dynamically adjust a network parameter of each network layer of the at least one network layer based on the image feature information.

16 . The apparatus according to claim 14 , wherein the network parameter comprises at least one of a convolution kernel parameter or an attention mechanism parameter, and the attention mechanism parameter is used to determine information that is in the image feature information and whose correlation with the point cloud data is greater than a correlation threshold as valid information for adjusting the point cloud feature recognition subnetwork.

17 . The apparatus according to claim 15 , wherein the operations comprise:

obtaining output data of the at least one network layer; and

determining, based on the output data, adjustment effect data corresponding to the network parameter.

18 . The apparatus according to claim 13 , wherein the semantic point cloud feature recognition network and the target detection network are obtained through training in the following manner:

obtaining a plurality of semantic point cloud training samples, wherein the plurality of semantic point cloud training samples comprise a point cloud data sample, an image sample projected to the point cloud data sample, and the semantic information of the image sample, and the three-dimensional location information of the target is labeled in the plurality of semantic point cloud training samples;

constructing the semantic point cloud feature recognition network and the target detection network, wherein an output end of the semantic point cloud feature recognition network is connected to an input end of the target detection network;

separately inputting the plurality of semantic point cloud training samples to the semantic point cloud feature recognition network, and outputting a prediction result by using the target detection network; and

performing iterative adjustment on network parameters of the semantic point cloud feature recognition network and the target detection network based on a difference between the prediction result and the labeled three-dimensional location information of the target, until iteration meets a preset requirement.

19 . The apparatus according to claim 11 , wherein the operations comprise:

performing panoramic segmentation on the image to generate the semantic information of the image, wherein the semantic information comprises a panoramic segmentation image of the image, and the panoramic segmentation image comprises image regions, obtained through panoramic segmentation, of different objects and category information corresponding to the image regions.

20 . The apparatus according to claim 11 , wherein the image comprises a panoramic image.

21 . A computer program product comprising computer-executable instructions stored on a non-transitory computer-readable storage medium that, when executed by a processor, cause an apparatus to:

obtaining an image and point cloud data of a target environment;

obtaining semantic information of the image, wherein the semantic information comprises category information corresponding to pixels in the image; and

determining three-dimensional location information of a target in the target environment based on the point cloud data, the image, and the semantic information of the image.