Back to results
US 12,754,307 B2Granted
Radar point cloud-based posture determination system
AssigneeSmart Radar System, Inc.
InventorsSoung On Kim, Myung Ku Kang
Life of this patent
Estimating the term…
Abstract
Provided is a radio detection and ranging (radar) point cloud-based posture determination system, which clusters a point cloud acquired using radar to generate a cluster, calculates predefined features of the cluster, and determines a posture of a target through a deep learning model that has trained the calculated features.
Claims14
A radio detection and ranging (radar) point cloud-based posture determination system comprising:
- a memory;
- and
- a processor, wherein the processor, when executing program instructions stored in the memory, is configured to perform:
- a data collecting operation of acquiring three-dimensional (3D ) point cloud data for a target using radar;
- a clustering operation of clustering the 3D point cloud data to generate a point cluster of the target;
- a feature calculating operation of calculating predefined statistical features on the basis of the generated point cluster;
- and
- a posture determining operation of determining a posture of the target on the basis of the calculated statistical features of the point cluster through a deep learning model trained to classify the posture of the target on the basis of the predefined statistical features of the point cluster,
- wherein the predefined statistical features include:
- (i) spatial distribution features including maximum values and minimum values of coordinates of points included in the point cluster in each axis direction of a 3D space, a height calculated as a difference between a maximum z-coordinate and a minimum z-coordinate of the point cluster, a width calculated as a difference between a maximum x-coordinate and a minimum x-coordinate of the point cluster, mean values of coordinates of the points in each axis direction, and variance values of the coordinates in each axis direction;
- (ii) shape ratio features including ratios between the spatial distribution features:
- (iii) temporal ratio features including a ratio of the height of the point cluster in a current frame to a maximum z-coordinate value or a minimum z-coordinate value of the point cluster in at least one previous frame;
- and
- (iv) motion features including a doppler average value obtained by dividing a sum of doppler values of the points by a total number of points in the point cluster;
- wherein the deep learning model classifies the posture of the target by using, as input data, a feature map generated from the calculated statistical features, without using the point cluster of the target, and
- wherein the statistical features from which the feature map for a current frame is generated include features that are computed based on data of the point cluster from at least one previous frame.
13 more elements
A method of determining a posture of a target, the method comprising:
- acquiring three-dimensional (3D ) point cloud data for the target using radar;
- clustering the 3D point cloud data to generate a point cluster of the target;
- calculating predefined statistical features on the basis of the generated point cluster;
- and
- determining a posture of the target on the basis of the calculated statistical features of the point cluster through a deep learning model trained to classify the posture of the target on the basis of the predefined statistical features of the point cluster,
- wherein the predefined statistical features include:
- (i) spatial distribution features including maximum values and minimum values of coordinates of points included in the point cluster in each axis direction of a 3D space, a height calculated as a difference between a maximum z-coordinate and a minimum z-coordinate of the point cluster, a width calculated as a difference between a maximum x-coordinate and a minimum x-coordinate of the point cluster, mean values of coordinates of the points in each axis direction, and variance values of the coordinates in each axis direction;
- (ii) shape ratio features including ratios between the spatial distribution features;
- (iii) temporal ratio features including a ratio of the height of the point cluster in a current frame to a maximum z-coordinate value or a minimum z-coordinate value of the point cluster in at least one previous frame;
- and
- (iv) motion features including a doppler average value obtained by dividing a sum of doppler values of the points by a total number of points in the point cluster;
- wherein the deep learning model classifies the posture of the target by using, as input data, a feature map generated from the calculated statistical features, without using the point cluster of the target, and
- wherein the statistical features from which the feature map for a current frame is generated include features that are computed based on data of the point cluster from at least one previous frame.
10 more elements
Dependencies are inferred from the claim text.