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.
- Every activity has a particular pose associated with it.
- Total 4 activities are consider for the scope of this work:
- Slap
- Kick
- Shoot
- Normal
- A comparative analysis of existing classifier to suite the data set.
git lfs fetch --all
Python (ver >= 3.4)
Numpy
Sklearn
OpenCv
- 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
- Z. Cao has proposed mutli-person pose estimation with using CNN
- Two branches - One for body part location and other for affinity between them.
- 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.
- Simple classification algorithms such as KNN, Decision Tree and Naive Bayes can be trained and used for classification.
- Performance of classifier of great importance
- Cross-validation is used for better evaluation of classifiers.
- 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)
- For further analysis KNN & DTree are selected
Click here to go the detailed report.