Smart Camera Development Using .NET MAUI, STM32H7, ESP32-S3 & LoRa Mesh

In this project case study, we are going to discuss how we at Adequate Infosoft designed and developed a high-performance, dual-connectivity AI-enabled smart camera system, showcasing our technical expertise in AI, IoT, embedded systems, wireless communication, and smart surveillance solution development.

Our design process encompassed developing customized hardware that supports Edge AI processing, developing a multi-protocol communication stack (SIM, Wi-Fi, and LoRa Mesh), and developing a cross-platform control application using .NET MAUI.

We recently built, tested, and released a custom AI-based smart camera system that will give businesses the ability to monitor remote locations such as factories, farms, construction sites, and remote security areas.

As part of the development process, we also established three independent connectivity options: built-in Wi-Fi, cellular (SIM-based 4G LTE), and LoRa-based mesh networking, allowing continuous communication with the camera, even from remote or difficult locations.

To facilitate the operation of this camera system, we created a cross-platform companion application using .NET MAUI (.NET 8), application enables users to control their devices across Android, iOS, Windows, and macOS devices without losing functionality.

The overall system achieved an edge detection accuracy rate of 98.7 percent, an alert latency rate of less than two seconds for both Wi-Fi and cellular connections, and eight seconds for LoRa mesh (up to five hops). The camera's battery will last up to 45 days when powered by a 10,000 mAh solar battery in low duty mode.

Locating "Dead Zones" due to limited access to both cellular and internet connections created a challenge for our client, who wanted to create a surveillance system in their large industrial complex and be able to monitor the system via mobile devices over the internet regardless of how many dead zones existed. To do this, the system needed to:

• Perform real-time AI (Human/Vehicle) detection without needing the Cloud.
• Use LoRa to maintain a mesh network so that alerts could then be relayed from each sensor back to a central gateway.
• Provide a mobile-friendly experience for monitoring the surveillance and configuring the system's hardware.

The hardware architecture and components were engineered to allow for the fastest AI inference possible, while also providing low-power, long-distance messaging capabilities. The hardware architecture consists of several components that can be incorporated into one custom PCB design:

Core Controlled:

The core controlling of the hardware is comprised of an STM32H7 Microcontroller in conjunction with either an NVIDIA Jetson Nano or Espressif S3 depending on the type of AI workloads being processed. This core controller will be used primarily for processing images.

Camera:

The OV5640 Optical Image Sensor has a resolution of 5 megapixels; however, it has an 8-bit color depth, making it a favorable choice for lower-light conditions.

Connectivity Suite:

• Simcom SIM7600: Generates a 4G/LTE backhaul network when Wi-Fi is not available.
• Espressif Wi-Fi and Bluetooth Module: Used for local high-speed video streaming and mobile pairing using cellular communications.
• Semtech SX1262 Low Long Range Transceiver: Provides the ability to create a mesh network of nodes that can communicate over 15 kilometers (or more) without obstruction (trees, buildings, etc.).

Power Supply:

A fully integrated lithium-ion battery pack that is equipped with a solar power header for remote off-grid power supply.

The main technical challenge of the LoRa Mesh implementation was using cameras that acted as "nodes" in areas with no 4G signal.

• Node behavior: Camera A will send out a lightweight encrypted packet using LoRa to notify of an intrusion detected by the camera if the camera does not have a SIM signal.
• Hopping: Camera B will receive this packet from Camera A (if it is within range) and will "hop" that packet to the Central Gateway.
• Gateway to Cloud: The Central Gateway has high gain antennas connected through an Ethernet/Fiber backhaul, and uses this network connection to send the alert for this intrusion to our Cloud Server.

We utilized TensorFlow Lite for Microcontrollers to run a pruned YOLOv8 model directly on the hardware.

• The Workflow: The camera captures a frame => The MCU/NPU runs inference => If "Human" confidence > 85% a trigger is sent.
• Benefit: This reduces data transmission costs by 95%, as the camera only "wakes up" the cellular module to upload video when a verified event occurs.

For the user interface, we chose .NET MAUI to maintain a single C# codebase for Android, iOS, and Windows.

Real-time Streaming:

Using the CommunityToolkit.Maui CameraView, we built a low-latency dashboard for live MJPEG streams.

Hardware Management:

The app communicates with the camera via Bluetooth LE (BLE) for initial setup (setting Wi-Fi credentials or LoRa frequency bands).

Smart Alerts:

Integrated Firebase Cloud Messaging (FCM) to deliver instant push notifications when the AI detects a threat.

Data Visualization:

We used Syncfusion MAUI Charts to show 24-hour activity heatmaps, helping users identify high-risk time slots.

This milestone was achieved through working very closely together with both our mobile (.NET) and firmware (C/C++) teams. Our efforts led to the following accomplishments:

Self-healing network:

A camera node failure causes the LoRa Mesh to automatically reroute the data to the next Camera Node.

Unified control:

Users can access the same application for local Wi-Fi monitoring and remote 4G monitoring without any hassle switching between locations using continuity of operation, including (but not limited to) switching between operations locally and remotely.

Summary of the Tech Stack

Although the complexity of the system is high, the solution engineering and delivery occurred on a single transaction from conception through delivery within the 6-month time frame (1.5 months for research/hardware designs, 2.0 months for firmware/AI model development, 1.5 months for the MAUI Application & Cloud Back-end, and 1 month for field testing and optimization).

This project illustrates Adequate Infosoft's ability to bring together modern cross-platform software engineering capabilities with complex hardware engineering to create a security solution that operates in situations where traditional security systems will not perform adequately.

Live App