The problem trinetra solves

Traditional security systems often struggle with being slow, unreliable, or resource-intensive. Camera-only solutions frequently generate false alarms due to animals, shadows, or lighting changes, while sensor-only systems lack the contextual intelligence needed to confirm real threats. Additionally, many modern systems rely heavily on cloud processing, introducing latency that makes real-time response difficult. Continuous video streaming further increases bandwidth usage and power consumption, making such systems inefficient and costly, especially in remote or resource-constrained environments.

Trinetra AI addresses these challenges by introducing a multi-layered intelligent IoT intrusion detection system that combines edge sensing, sensor fusion, and AI-based visual validation. It processes data at different levels of the pipeline, enabling faster and more accurate decision-making. At the edge, the ESP32 filters noise using sensor voting and environmental checks, reducing unnecessary alerts. The Raspberry Pi then performs multi-sensor fusion to improve reliability, while the final AI layer uses deep learning to visually confirm whether a detected object is truly a human, significantly reducing false positives.

This system can be used in smart home security, industrial monitoring, restricted zones, campuses, and even remote or border surveillance where connectivity is limited. By operating in an event-driven manner, it activates high-level processing only when needed, thereby saving bandwidth and power. Overall, Trinetra AI makes security systems smarter, faster, and more efficient by reducing false alarms, enabling real-time responses, and minimizing resource usage, offering a reliable alternative to traditional surveillance solutions.

Challenges we ran into

1. Camera module failure during hackathon → quickly pivoted to using a mobile phone as a live camera feed, ensuring zero downtime in visual detection

2. Difficulty in sourcing domain-specific audio datasets → conducted extensive research, combined open-source datasets (GitHub + online sources) and generated synthetic data to fill gaps

3. Training specialized sound detection models → optimized dataset quality and preprocessing to achieve reliable model performance despite limited data

4. Raspberry Pi OS booting issues → faced repeated boot failures, resolved through continuous reflashing and boot debugging

5. Compatibility issues with Windows during Pi setup → identified environment limitation, switched to macOS where setup and flashing worked seamlessly