Target object position prediction and motion tracking
A computer-implemented method for target object position prediction includes receiving, via an RGB camera a plurality of images depicting one or more persons positioned on a floor. A plurality of person location labels is assigned to each image indicating where the one or more persons are located relative to the floor. A foot position (FP) classifier is trained to classify the images into the person location labels, wherein the FP classifier is configured according to a multi-layer architecture and the training results in determination of a plurality of weights for connecting layers in the multi-layer architecture. A deployment of the FP classifier is created based on the multi-layer architecture, the plurality of weights, and the plurality of person location labels.
1. A computer-implemented method for target object position prediction, the method comprising:
receiving, via a RGB camera a plurality of images depicting one or more persons positioned on a floor;
assigning one of a plurality of person location labels to each image indicating where the one or more persons are located relative to the floor;
training a foot position (FP) classifier to classify the images into the person location labels, wherein the FP classifier is configured according to a multi-layer architecture and the training results in determination of a plurality of weights for connecting layers in the multi-layer architecture;
creating a deployment of the FP classifier based on the multi-layer architecture, the plurality of weights, and the plurality of person location labels.
2. The method of claim 1 , wherein the FP classifier is trained using coordinates of the RGB camera.
3. The method of claim 2 , wherein the FP classifier is further trained using a plurality of keypoints corresponding to the one or more persons and lines indicating connections between connected keypoints.
4. The method of claim 3 , wherein the keypoints are generated by a pose estimation (PE) classifier based on the images.
5. The method of claim 1 , further comprising:
determining a subset of the images where at least one of the persons is not occluded by any other objects in the images,
wherein the FP classifier is only trained using the subset of images.
6. The method of claim 1 , wherein the deployment of the FP classifier comprises one or more files describing (i) the multi-layer architecture, (ii) the plurality of weights, (iii) the plurality of person location labels.
7. The method of claim 1 , wherein the FP classifier is a convolutional neural network (CNN).
8. A computer-implemented method for target object position prediction, the method comprising:
receiving, via an RGB camera, a plurality of images depicting a person positioned on a floor;
applying a trained pose estimation (PE) model to the images to determine a plurality of keypoints associated with the person;
determining the person's location relative to the floor by applying a trained foot position (FP) classifier to inputs comprising the keypoints and coordinates specifying a location of the RGB camera; and
providing a visualization of each person's location relative to the floor.
9. The method of claim 8 , further comprising:
identifying a plurality of connected keypoint pairs,
for each connected keypoint pair, generating a line,
wherein the inputs to the FP classifier comprise the lines generated for the connected keypoint pairs.
10. The method of claim 8 , further comprising:
analyzing the plurality of images to confirm that the person is not occluded prior to applying the FP classifier.
11. The method of claim 8 , wherein the FP classifier projects a center of gravity of the person downwards onto a plane of the floor to determine the person's location relative to the floor.
12. The method of claim 8 , wherein the FP classifier is a convolutional neural network (CNN).
13. The method of claim 12 , wherein the PE model is a CNN.
14. The method of claim 8 , wherein the visualization comprises an indication of the person on a building floor plan.
15. The method of claim 14 , wherein the visualization further comprise a depiction of the keypoints overlaid on the person in the images.
16. A system for target object position prediction, the system comprising:
a pose estimation (PE) model trained to determine a plurality of keypoints associated with a person using images acquired with an RGB camera;
a trained foot position (FP) classifier trained to determine the person's location relative to a floor in the images based on the keypoints and coordinates specifying a location of the RGB camera; and
a visualization model configured to provide a visualization of the person's location relative to the floor on one or more displays.
17. The system of claim 16 , wherein the trained FP classifier is a convolutional neural network (CNN).
18. The system of claim 16 , further comprising:
an occlusion detection module configured to analyze images to confirm that the person is not occluded prior to applying the FP classifier.
19. The system of claim 16 , wherein the visualization comprises an indication of the person on a building floor plan.
20. The system of claim 19 , wherein the visualization further comprise a depiction of the keypoints overlaid on the person in the images.