Pose estimation analysis with monocular camera

1 . A system configured for implementing a temporal-based two-dimensional (2D) human pose estimation method for golf swing analysis using temporal information to improve pose estimation for changing positions and partially self-occluded key points, the system comprising:

one or more hardware processors configured by machine-readable instructions to:

select a first image as a representative image from a video sequence comprising a plurality of temporally adjacent frames to the first image;

determine a bounding box to incorporate with the first image of the video sequence received from a user device;

initiate a temporal 2D inference process on the first image to generate a second 2D image, wherein the temporal 2D inference process aggregates features across temporally adjacent frames is executed on the first image within the bounding box and the temporal 2D inference process generates a set of confidence scores for one or more key points of the first image within the bounding box;

provide a second 2D image and the set of confidence scores for at least one key point of the one or more key points of the first image to a three-dimensional (3D) inference process and Perspective-n-Point (PnP) process;

generate a 3D image using output from the 3D inference process, wherein the 3D image is modified based a distance value between a user and a camera of the user device that is determined by the PnP process; and

provide the 3D image to the user device.

2 . The system of claim 1 , wherein the one or more hardware processors configured by the machine-readable instructions are further configured to:

receive one or more input images from the user device, wherein the one or more input images are each associated with an original resolution;

for each of the one or more input images, reduce the original resolution of the one or more input images to a reduced resolution;

initiate a feature extraction process on the one or more input images stored at the reduced resolution to generate one or more feature maps of the one or more input images;

apply a heatmap regressor to each of the one or more feature maps; and

associate locations of key points with the one or more input images at the original resolution.

3 . The system of claim 1 , wherein the PNP process estimates a relative pose of the user between the user and the user device given a set of correspondences between 3D points and their projections on an image plane.

4 . The system of claim 1 , wherein the bounding box is programmatically placed around the key points of the first image by predicting locations of a golf club and the user.

5 . The system of claim 1 , wherein the first image received from the user device is included with a series of three images and a second or a third image from the series of three images.

6 . The system of claim 1 , wherein the one or more hardware processors configured by the machine-readable instructions are further configured to: train a machine learning (ML) model to recognize people and sports equipment based on images that show these objects and do not show these objects.

7 . The system of claim 1 , wherein the one or more hardware processors configured by the machine-readable instructions are further configured to:

train a machine learning (ML) model to programmatically learn differences between pixel data and objects in the first image to recognize people and golf clubs in other images.

8 . The system of claim 1 , wherein the one or more hardware processors configured by the machine-readable instructions are further configured to:

use a trained machine learning (ML) model to predict a location of a person and a golf club in the first image and place the bounding box around the person and the golf club.

9 . The system of claim 8 , wherein the trained ML model includes a reduced number of layers in a convolution of the trained ML model than traditional ML models.

10 . The system of claim 1 , wherein the temporal 2D inference process is initiated to determine X and Y coordinates with a confidence score for each point.

11 . The system of claim 1 , wherein the one or more hardware processors configured by the machine-readable instructions are further configured to:

apply a line segment detection algorithm, as part of the temporal 2D inference process, to a portion of the first image within the bounding box; and

generate, using the line segment detection algorithm, output comprising a set of unconnected line segments pointing in various directions.

12 . The system of claim 11 , wherein generation of the set of unconnected line segments depend only on pixel information, rather than connecting lines in the first image.

13 . The system of claim 11 , wherein the set of confidence scores for one or more key points of the first image within the bounding box correspond with a number that represents a likelihood that output of a line segment detection algorithm within the bounding box is correct.