Nirikshan (Supervision)

Real Time Video Analytics Pipeline using object detection and tracking

*NOTE --> ByteTracker uses python-package lap, which wont work for python-version > 3.10*

Nirikshan aims to provide video analytics on video sources like RTSP stream or video file using deep learning models.

• Video Source
  • Video file
  • RTSP Stream
• DNN Models
  • object detection: yolo-V3/V4
  • Pose Estimation: MediaPipe
  • Face Detection
• Object Tracker: ByteTracker
• OpenCV for Video Processing and DNN for Inference
• NVIDIA Triton for Inference
• User Interface
• Python

Analytics

• Class: Person
  • Count
  • Emotion NotImplemented
  • Age Category NotImplemented
    • Young AdultsNotImplemented
    • Middle AgedNotImplemented
    • Older AdultsNotImplemented
  • Action
    • Smoking
    • Fighting
• Class: Vehicle
  • Count
  • Type (Car, Bus, Bike)
  • Color
  • Brand
  • LPR
• Class: Animal
  • Count
  • Species

How to run

1. Install dependencies using conda, use conda env file myenv.yml
2. Download models from below links
   • YOLO Models
   • Yunet-face
3. Make sure conda env is activated.
4. To Run application, Use command make run FPATH=data/palace.mp4

FUTURE SCOPE (Items in the list are in consideration, not finalized though)

• Action recognition in videos.
• Support for multiple video sources (IP Cameras, Local Files)
• Dockerize Analytics
• GPU Support for faster inference
• User Interface for visualizing analytics results

References/citations

• yolov3: Radmon et al. "YOLOv3: An Incremental Improvement"
• yolov4: Bochkovskiy et al. "YOLOv4: Optimal Speed and Accuracy of Object Detection"
• ByteTrack: Zhang,Yifu et al. Multi-Object Tracking by Associating Every Detection Box.
• YuNet: https://github.com/opencv/opencv_zoo/tree/main/models/face_detection_yunet
• PaddleDetection: https://github.com/PaddlePaddle/PaddleDetection/blob/develop/deploy/pipeline/README_en.md