Hi everyone,
I’m planning a project to enhance my smart home by adding face recognition capabilities to trigger automations in Home Assistant. My goal is to use a second Raspberry Pi 5 (16GB RAM) along with the Raspberry Pi AI Kit (26 TOPS NPU) to run custom AI models for facial recognition, primarily to identify family members and trigger relevant automations (e.g., turning on lights, unlocking doors, sending notifications, etc.).
Setup details:
• The video feed will come from 4-5 cameras via RTSP streams.
• I want a solution that leverages the AI acceleration of the Raspberry Pi AI Kit.
• The system should integrate well with Home Assistant, preferably via MQTT or other efficient methods.
• I’d like to train and add custom face recognition models.
• Low latency is important to ensure near-instant automation triggers.
I’ve researched several possible approaches but I’m a bit overwhelmed by the options available. Some solutions I’ve come across include:
• Frigate + MQTT – Seems to be a popular choice for object detection and RTSP streams, but how well does it handle face recognition with custom models?
• DOODS (Dedicated Open Object Detection Service) – Lightweight option, but unsure about its accuracy and performance on a Raspberry Pi.
• TensorFlow – Powerful but may be too resource-intensive; is it practical for my hardware?
• DeepStack – Easy to set up, but might have limitations with custom models and multi-camera support.
I’m looking for advice from anyone who has experience with a similar setup or can recommend the most efficient and practical software stack to achieve this. Any insights on the best way to process multiple RTSP feeds while maintaining good performance would be highly appreciated.
Thanks in advance for your suggestions and advice!