Abnormal Event Detection | ConvLSTM Autoencoder

Recorded analysis

Review a full clip with frame-level context

Upload a video and inspect anomaly scores, flagged frames, and summaries.

MP4, MOV, AVI Up to 100 MB Remote inference

Drop a video here or click to browse

Best results with stable lighting and fixed cameras.

Recommended 720p or 1080p, steady motion

Tip Raise threshold to reduce false alerts.

Anomaly threshold Higher means fewer alerts. Try 0.05-0.1 in calm scenes.

Quick thresholds

Max frames Cap frames to control processing time and memory.

Processing happens on the inference service with your selected parameters.

Live monitoring

Monitor a camera feed with instant alerts

Start a local camera, tune thresholds, and overlay detected objects.

10 FPS target Face-aware suppression

How does the live feed work? How do I test it?

How it works: Your browser captures video from your camera and sends frames to the server every 100 ms. The server runs (1) the ConvLSTM anomaly model, (2) face detection, and (3) object detection (YOLO). If the anomaly score is above the threshold and no face is detected, an alert is shown (so your face in view will not trigger alerts). Detected objects (person, car, etc.) are drawn on the video.

Camera not starting? Browsers only allow camera access on HTTPS or localhost. If you opened the app via an IP address (e.g. 192.168.x.x), use http://localhost:5000 on the same machine instead, or serve the app over HTTPS.

How to test: (1) Click Start Camera and allow camera access. (2) Look at the camera - alerts are suppressed when a face is detected. (3) Move out of frame or show unusual motion to see the score change. (4) Use a high threshold (e.g. 0.05-0.1) to reduce false alarms.

Normal Score: --

Anomaly threshold Alerts are off when a face is detected - your face will not trigger alerts. Show object detection (boxes on video)

Privacy note: Alerts are automatically suppressed when a face is detected in the frame.

FPS: --