CamAi takes an AI first approach to monitoring and description of a video stream. Traditional open source packages like ZoneMinder and others largely rely on motion detection and handcrafted "zones" to try to reduce false detections. Modern Cameras support much higher resolutions (FHD, 4MP, 5MP, 4K etc) and this requires consistently low latency handling of the video streams, which the older systems seem to be having issues coping with.
CamAi tries to take advantange of the 'many cpu cores' and is heavily threaded to ensure reliable high resolution video stream processing of 'many' sources. It also uses more modern approaches like Instance Segmentation, Face Detection and Face Recognition to reduce the false alarm rates. While this requires a GPU/s to get reasonable performance, inference engines bundled with traditional cpu cores are the future. It takes a data science approach to enable learning over time what might be normal or otherwise and get closer to a human in taking actions. Some of these features are available in the "cloud" cameras like Ring tm;, Nest tm;, which requires the video streams to be sent to 'their' cloud, and come with many involuntary 'sharing' features. CamAi can do this all within the network, with or without any external connectivity for the camera streams.
It is certainly possible to try to 'graft' these features to existing systems. However it would have resulted in a mish mash of languages and technologies (perl/c++/windows only) or only partially open source. CamAI aims to have a more modern, maintainable and extensible code base that is fully be open source, and cross platform (over time) and modern, without having to jettison control and visibility to 'cloud' cameras.
The development has been primarily on a ubuntu 18.04 base.
- Install Ubuntu Packages in ubuntu_deps.txt.
- bash ./ubuntu-deps.txt
- Install Nvidia CUDA for tensorflow acceleration
Any recent compatible combination of Cuda, CuDNN and Tensorflow should work
You can also run this without an NVIDIA GPU on a many core system, especially if the camera resolutions are 1080p or lesser.
- Install Pip3 packages in requirements.txt
- Optional but Recommended: Create a virtual env for camai
- mkdir camai # If you do not have virtualenv already, or want to use a more current version pip3 install virtualenv --user virtualenv -p python3.7 venv37
- Unpack CamAi
- tar xvfz - camai.tgz (from tar bundle) pip3 install ./camai.whl (from a local wheel, pypi has size limits)
- Install required python packages
- pip3 install -r ./CamAi/requirements.txt
- Quick Run
Discover onvif compatible cameras on your network and generate a configuration file
./CamAi/camaicli.py discover or if installed from a wheel python3.7 ~/venv37/lib/python3.7/site-packages/CamAi/camaicli.py discover --config ./aicam.toml
Start monitoring with ./CamAi/camaicli.py monitor --config ./aicam.toml or if installed from a wheel python3.7 ~/venv37/lib/python3.7/site-packages/CamAi/camaicli.py monitor
If you want to specify a configuration file in a different location you can add --config config.toml to the above python3.7 ~/venv37/lib/python3.7/site-packages/CamAi/camaicli.py monitor
This should start logging videos to storage directories you specified in the config file. The default location is ./video You cannot open the video being currently logged till it's rotated. Default rotation is 30 minutes at hourly boundaries, so you should be able to open it after next hours starts. Alert images and snippet videos are generated in realtime and are located at the base directory for viewing.
Below is an example configuration file that the camcli.py --discover option generates. The configuration file can be hand crafted. There are three main sections.
- Cameras
- There has to be one or more camera sections. This section contains camera specific configuration and policy information. The camera section consists of a main section and a sub section. The sub section is used to specify one or more objects of interest for this camera and the object specific settings.
- Email Section
- This contains one Email Sender and zero or more Email Recepient sections
- Manager Section
This section has CamAi runtime defaults. If the user doesn't specify a per camera directory, the manager sections basedir variable is used as the parent for each camera.:
[[camera]] hostname = "192.168.2.10" <--- Largely ceremonial (Not mandatory) port = 8000 <--- Onvif port (Not mandatory) username = "admin" <--- Not used yet, earlier versions used this to construct the URL password = "blahblah" <--- Not used yet, earlier versions used this to construct the URL name = "Garage Walkway" <--- Make this human consumable as it shows up in alerts/logs etc (Required) url = "rtsp://admin:[email protected]:554/h264Preview_01_main" (Required) fps = 22 <--- Gets used, use a sane number (Recommended) snapshoturl = "http://192.168.1.10:80/onvifsnapshot/snapshot_channel01.jpg" <--- (Not used yet) readbuffer = 78643200 <--- 16 x resolution is generally recommended (Required) resolution = "2560x1920" <--- Not mandatory enableptz = false <--- Not used yet subdir = "./videos/Garage Walkway" <--- Recommended, manager sections basedir + name is used otherwise mode = "detect_and_record_timelapse" <--- Required rotation = 0 <--- Required if image is not oriented correctly watchdogenable = true <--- Only set to false if camera is actually just a video file watchdogtimer = 60 <--- Default is good annotation = false <--- Leave to false other than when image masking figuring out false detects, intensive overhead showlivevideo = false <--- Not used yet, older versions us facedetection = false <--- Hmm, this needs to deprecated as 'instancedetection' under objects of interest is the right maxreadqueue = 32 <--- Tweak per your memory availability, memory consumed is this times resolution detectionrate = 25 <--- Detection frequency, these many frames are skipped between detections detecttrackseconds = 30 <--- Automatically track this many seconds if object of interest is detected deletepolicyenabled = true <--- Enable automatic storage management deletethreshold = 60 <--- Start deleting when overall disk usage reaches this percentage deleteeventsthreshold = 80 <--- Start deleting videos with events when overall disk usage reaches this percentage deleteafterdays = 7 <--- Only delete after this many days log_object_timeseries = true <--- Leave at true, used for instance detection and reducing redundant notifications ffmpeg_options = "" <--- Video decoder / transport to user, defaults to OS s/w (minorchance for a race condition causing incorrect decoder being used if using threaded observer. Example Nvidia H264 H/W decoder + TCP transport: "video_codec;h264_cuvid|rtsp_transport;tcp" [camera.objects_of_interest.person] <--- Sub section for an object of interest, in this case a 'person' object detection_threshold = 0.98 <--- How confident the model should be that this object was detected instance_match_threshold = 75 <--- How confident the model should be that this is the known person based on face recognition instance_watch_timerange_start = 8 <--- Suppress alerts if the same object is seen within the last 8 minutes instance_watch_timerange_end = 2 <--- Works with above, when to start looking if the object was last seen, default is 2 mins ago instance_watch_continuous_mode = false <--- The object has to be continually present within the watch timerange otherwise alerts will be triggered instancedetection = false <--- Face recognition if it's a person object, (licenses for vehicles will also use this) notify_startup_wait = 5 <--- How long to wait after startup before notifying, otherwise it will report existing objects [camera.notifications] local_audio_notification = true <-- Enable/Disable local audio alerts for this camera email_notification = true <-- Enable/Disable email alerts for this camera retain_clips = true <-- Retain or Delete the alert video clips / images after notification [camera.objects_of_interest.car] detection_threshold = 0.85 instance_match_threshold = 65 instance_watch_timerange_start = 6 instance_watch_timerange_end = 2 instance_watch_continuous_mode = false instancedetection = false notify_startup_wait = 5 [camera.objects_of_interest.motorcycle] detection_threshold = 0.98 instance_match_threshold = 75 instance_watch_timerange_start = 8 instance_watch_timerange_end = 2 instance_watch_continuous_mode = false instancedetection = false notify_startup_wait = 5 [camera.objects_of_interest.truck] detection_threshold = 0.85 instance_match_threshold = 65 instance_watch_timerange_start = 6 instance_watch_timerange_end = 2 instance_watch_continuous_mode = false instancedetection = false notify_startup_wait = 5 [camera.objects_of_interest.bus] detection_threshold = 0.98 instance_match_threshold = 75 instance_watch_timerange_start = 8 instance_watch_timerange_end = 2 instance_watch_continuous_mode = false instancedetection = false notify_startup_wait = 5 [camera.objects_of_interest.bicycle] detection_threshold = 0.98 instance_match_threshold = 75 instance_watch_timerange_start = 8 instance_watch_timerange_end = 2 instance_watch_continuous_mode = false instancedetection = false notify_startup_wait = 5 [["email recepient"]] <--- Can have any number of recipients name = "First Last" <--- Recipient Name email_address = "[email protected]" <--- Recipient Email attachment_preference = 'image_only' <--- One of image_and_video(default), image_only, video_only or no_attachments ["email sender"] sender_email = "[email protected]" <--- Email address from which the alerts are sent sender_login = "[email protected]" <--- Login to use to authenticate with the smtp server smtp_server = "mail.example.com" <--- Your email providers SMTP smtp_server_port = 465 <--- SSL SMTP Server Port login_required = true <--- Only authenticated SMTP is supported use_SSL = true <--- Only SSL SMTP is supported sender_secret = "blahblahpassword" <--- The password ["manager options"] basedir = "./videos/" <--- Default storage location if not overriden on a per camera basis numdetectors = 1 <--- Number of detector processes, if more than 1, multiprocessing_detector and multiprocessing_observer has to be true pipelineresize = true <--- Image resizing for detection, doing it per camera (vs in detector) reduces latency singledetectorqueue = true <--- Use a single queue for all cameras vs one queue for all cameras, single is cheaper defaultmaxqueue = 32 <--- Default queue size, per camera override exists multiprocessing_observer = false <--- Leave these settings as is for now multiprocessing_detector = false multiprocessing_reader = false multiprocessing_writer = false multiprocessing_notifier = true <--- Has to be true if face recognition is desired multiprocessing_viewer = false multiprocessing_trainer = true
- Choosing the Camera Operating Mode
CamAi supports four operating modes. The mode is configurable on a per camera basis. The default mode generated by the
- detect_and_record_timelapse
- Detection is on for the camera, and frames are recorded in a timelapse fashion at sample interval specified by 'detectionrate'. If any objects of interest are observed, the recording becomes continuous the number of seconds specified by 'detecttrackseconds' for that camera. If facedetection/instancedetection is enabled, the frames are further processed to scan for known instances before generating alerts.
- detect_and_record_everything
- This mode works similar to detect_and_record_timelapse, except that all frames are recorded whether or not an object of interest is detected or otherwise.
- detect_only
- This mode works similar to detect_and_record_timelapse, except that no video is recorded (other than short alert clips). Alerting etc. works exactly the same.
- record_only
- This mode turns off any detection and records every frame.
- Setting up Face Recognition
Create a known faces directory. Follow the sample directory heirarchy. Put this under camai directory for now. Pictures in standard formats like jpg/gif/png/bmp should work. Each picture should only have the face of the person whose name is the parent directory.:
known --->(Parent Directory) people --->(People Directory) │ ├── barack obama ------> (Directory is named after the person) │ └── barack.jpg │ └── vacation1.jpg │ └── dude3.jpg └── Da Wife │ └── Wife1.png │ └── Spouse2.png │ └── IMG2019_3838383.JPEG │ └── BossLady.png └── Da Hubby │ └── dameek1.png │ └── papertiger.png └── Da Son └── slacker.jpg └── gamerboy.jpg └── prodigalreturns.jpg- Alerts
The following types of alerts are supported.
- Verbal
- When a new object of interest is detected, it is announced locally on the computer thats running camai.
