martlettt / human-pose-detection

Implementation and quantitative analysis of different classifier on top of pose data from separate model.

Human-Pose-Detection

Objective

According to NCRB, 2.97 million cases of crime recorded in year 2018. The project propose a solution for remote monitoring and analysis, suitable an aerial vehicle - Suspicious activity detection through video analysis, primarily for human pose detection using visual features.

Project information

• Every activity has a particular pose associated with it.
• Total 4 activities are consider for the scope of this work:
  1. Slap
  2. Kick
  3. Shoot
  4. Normal
• A comparative analysis of existing classifier to suite the data set.

Code Execution Instructions

Command to download large files

git lfs fetch --all

Requirements

Python (ver >= 3.4)
Numpy
Sklearn
OpenCv

Steps

• Orientation Extraction on Images

python OpenPoseImage.py

• Training (Results stored in 'orient_train.csv')

python multi-person-train.py

• Classification on video (sample video: 'etc/d_fight.mp4') - Using Dtree/ KNN classifiers

python multi-person-classify_video_dtree.py -v video_path

python multi-person-classify_video_knn.py -v video_path

• Testing (Results stored in 'orient_test_result.csv')

python multi-person-classify_test_knn.py

Process Flow

Pose Estimation

• Z. Cao has proposed mutli-person pose estimation with using CNN
• Two branches - One for body part location and other for affinity between them.

Orientation Extraction

• Angle with 13 major pairs of body is considered such as Shoulder to Elbow and so on.
• Angles are inverted and w.r.t to horizontal axis.

Classification

• Simple classification algorithms such as KNN, Decision Tree and Naive Bayes can be trained and used for classification.

Analysis of Classification Techniques

• Performance of classifier of great importance
• Cross-validation is used for better evaluation of classifiers.

Results

Output of the proposed method

Pre-processing of Data

• Due to obstruction in the scenario many body parts will not get covered.
• A weak assumption that those body parts are vertically straight is made.(highlighted by yellow color)

Data set Statistics

Comparative Analysis results

KNN

More

• For further analysis KNN & DTree are selected

Video Demonstration

KNN (individual frames)

Click here to go the detailed report.