Modern compression techniques leverage a learned representation over an image distribution to more efficiently encode redundant spatial context for a sample. Video content is also generally redundant in time.

This work implements online training of convolutional autoencoders to learn a nonstationary image distribution in streaming video. Then, we perform anomaly detection by thresholding the autoencoder's reconstruction loss, which we regard as an anomaly score.

Adapt the included config.ini to suit your needs.

The demo uses OpenCV so you can configure the video sources with:

• a webcam index (0,1 or /dev/video0)
• an RTSP url
• other URIs recognized by cv2.VideoCapture() (.mp4 or .mkv containers)

To affect model recency bias, consider changing:

• the model learning rate (smaller rates --> less recency bias)
• the time window length (smaller windows --> greater recency bias)

To explore different network architectures, change the net_arch parameter. This should be a list of 3-tuples organized as:

• First entry - No. of Filters (Output dim)
• Second entry - Size of Convolutional Kernel (assumed square)
• Third entry - Size of Pooling (assumed square)

This repo requires opencv and tensorflow:

pip install -r requirements.txt

Set the config to reflect your video sources and model preferences before running:

python3 visual_anomaly_detection_demo.py

A network is instantiated for each video source according to preferences set in the config.ini.

Online training begins and normal/anomaly classes along with the reconstruction loss will be streamed to stdout according to the configured threshold (by default 5 sigma deviation from moving average).

• Hackster/NVIDIA - AI at the Edge Challenge Entry