AI based Human Body Parts Detection

Source: Deep Learning on Medium

AI based Human Body Parts Detection

Abstract: In this paper, we investigate deep learning methods that will automate the Human Part Detection of the human body. Most of the detection process currently rely on manual system. There is a need of AI-driven Human Part Detection of the human body so that use case related to Pose detection can be performed. This paper consists of all the details of four parts(Head, Torso, Hand and Full body) detection of human body and how each part can be validated by different Deep learning Model sequentially.

1.INTRODUCTION

AI based real-time Human Part Detection of the human body can be described as four human parts(Head, Torso, Hand and Full body) detection. We will explain this detection process by assuming that different parts of human parts is being captured on camera and then image is feed to Deep Learning Image Validator for the purpose of validation. Deep Learning Image Validator is being used to validate four parts of image and each part of human body will be validated by four different Deep learning Model:1. Validation of Head 2. Validation of Torso. 3. Validation of Hand 4.Validation of leg along with Head, Torso and Hand.

2.MOTIVATION:

Human Pose estimation is an important problem that has enjoyed the attention of the Computer Vision community for the past few decades. It is a crucial step towards understanding people in images and videos. First steps towards Human Pose detection is to detect Key Point of these four parts of the human. Simple Face Detection can also be performed by implementing Head detection logic of this paper.

3. DEVELOPMENT WORK:

Creating own dataset is very tedious tasks so we can use open source dataset available in the market. Dataset should consist human image or human parts along with proper annotation.

In this paper, We are using pre-trained Caffe model(pose_iter_440000.caffemodel) trained for Pose detection. It detects 18 key points of the body. These are Nose, Neck, RShoulder,LElbow,RWrist,LShoulder,LElbow,LWrist,RHip,RKnee,RAnkle,LHip,LKnee, LAnkle, REye,LEye, REar, LEar, Background.

4. SOLUTION :

We developed four different web application using Flask Framework to perform validation on set of images and single image.

If you want to validate set of images at a time then Validate images of the file can be selected in above screen. And result of that will look like this:

Each Validator can also be Visualized.We have implemented this using matplotlib library of python. Pass indicates that out of total images of the folder,12 have passed head validator and 4 have failed head validator.

All Validator can also be visualized together as well.

5. TECHNICAL REQUIREMENT:

Python can be used as programming language.

Jinja with Flask Framework Template can be used for building UI application

6. CONCLUSION:

1. Model performed with 95% accuracy on training data and 93% accurate on test data.

2. 50% of confidence score was taken for head part of detection.

3. Model is performing poorly on noisy image.

7.REFERENCES:

https://lmb.informatik.uni-freiburg.de/Publications/2016/OB16a/oliveira16icra.pdf

https://hypjudy.github.io/2017/05/04/pose-estimation/

https://github.com/yysijie/openpose