wjtscn / daywatch

Moving objects detection and recognition

Daywatch

The project main goal is to detect and recognize moving objects using outdoor camera video stream. Pretrained YOLOv3 model is used as recognizer. YOLOv3 is trained on COCO dataset. More details about COCO classes can be found here.

When one or more object is detected, the screenshot can be saved. It is possible to manage background classes, objects of such classes do not trigger screenshot. The YOLOv3 part of the project is based on original Kaggle notebook.

If the camera supports ONVIF protocol, it is possible to control camera by press and release mouse left button on selected areas of the screen. Both RTSP and MJPEG video streams are supported (for the latter use key -mjpg), which makes it possible to use cheap webcams plugged into router with OpenWRT and professional videosurveillance devices.

Dependencies

• Python 3.6.x
• Tensorflow 1.1x
• OpenCV 4.x
• NumPy
• onvif_zeep

Weights

Load YOLOv3 weights from https://pjreddie.com/media/files/yolov3.weights and use key -wf (or --weights-file) to provide path for weights file (default is ./yolov3.weights)

Getting started

Use key -h to read about all available options

daywatch.py -h

When focus on security feed window following hotkeys available:

• <space> manually save a screenshot
• m switch to multiscreen mode, which might be useful for tune motion detector parameters
• b show/hide background zones (when background objects are defined in json file)
• c in background mode: switch between background zones for specific classes
• q quit

Remote control

If camera provides ONVIF endpoint and ONVIF credentials are provided ( -oc or --onvif-credentials key), remote control is availabe based on continuous moves. If mouse right button is pressed at the left third part of the image, camera starts moving left, if it is pressed at the right third part of the image, camera start moving right, and similar for the upper and lower third parts of the image. These moves can be combined: if mouse right button is pressed near the corner of the image, camera start moving top-left, top-right, bottom-left or bottom-right, depending of the corner. Camera stops when mouse right button is released. Due to non-zero latency, camera can stop moving with some delay. While camera is moving, no motion detection and object recognition is performed.