| AID-R Assault Detected on Jetson Nano (4 FPS) | Normal Activity Detected |
|---|---|
 |
 |
AID-R is a low power, edge-compute platform utilizing the Jetson Nano and Keras/Tensorflow-based automated human activity recognition to provide smart cities with ad-hoc, video based neighborhood watch services to support assault/robbery-in-progress assessment, public safety, property protection, etc. The system can be implemented for example on forward- and reverse-view car cameras, with existing cctv networks or stand-alone mounted systems. Such a smart city adhoc video anaytic criminal activity monitor system addresses public privacy concerns by automating the detection and recognition of person/property assault by live monitoring video feeds, not recording video and eliminating human-in-the-loop involvement when normal activity is detected.
The repository contains:
- Code for training (train.py) and inferencing (predict_camera.py) on the Jetson Nano
- Example video clips to use for inferencing source
- A serialized model trained on the sports activity dataset
- A label pickle
- An install script for systemd startup service for the AIDR application
Usage python predict_camera.py --model model/activity.model --input example_clips/assault.mp4 --label-bin model/lb.pickle --output output/assault.avi --size 100
Boolean flags inside predict_camera.py can be adjusted to input from video, record new training/test data and to record inferenced video.
Jetson Nano and either csi or usb camera:
Example image of the stand-alone AIDR mounted in forward view camera position on vehicle dash:
- Python3
- Tensorflow GPU >= 1.13.1
- Keras > 2.3.1e
- Opencv > 4.1.0
***Dataset:
Data Download link in References section below.
Data Distribution Training set: 11524
Validation set: 2881
**Pretrained Model- Resnet-50
**Data Augmentations- The following data augmentation has been applied to increase the no of images in the training set:
Flip horizontal Lighting Zooming Warping
- Dataset collection from dash mounted standalone unit
- Improve power performance
Dataset: https://www.dropbox.com/s/0jp57lhs0y805ro/sports-type-classifier-data.7z?dl=0
Tutorial on Keras video classification: https://www.pyimagesearch.com/2019/07/15/video-classification-with-keras-and-deep-learning/
Configuring Jetson Nano: https://www.pyimagesearch.com/2019/05/06/getting-started-with-the-nvidia-jetson-nano/
Increasing swap memory and installing Jetcam: https://thenewstack.io/tutorial-configure-nvidia-jetson-nano-as-an-ai-testbed/
Install Opencv on Nano: https://pythops.com/post/compile-deeplearning-libraries-for-jetson-nano
Nano case: https://www.amazon.com/gp/product/B07ZNVH982/ref=ppx_yo_dt_b_asin_title_o03_s00?ie=UTF8&psc=1
CSI cam: https://www.amazon.com/gp/product/B07VFFRX4C/ref=ppx_yo_dt_b_asin_title_o03_s01?ie=UTF8&psc=1
USB cam: https://bluerobotics.com/store/sensors-sonars-cameras/cameras/cam-usb-low-light-r1/