InnoComm Helps Bring AI Integration to a Smart Care Device on the MediaTek Platform

The use of AI in remote and healthcare applications is evolving from basic video monitoring toward real-time risk detection and proactive alerts. Smart care devices must do more than provide visibility into on-site conditions. They must also identify falls, bed-exit events, and other behaviors that may put residents at risk, enabling caregivers to intervene promptly.

InnoComm helped a customer develop an AI-powered smart care device for remote and healthcare environments. The system can connect to caregiver beacons, resident wearables, emergency buttons, and other sensing devices. It provides real-time safety alerts, contextual event information, and privacy-protected video, helping care teams assess situations more quickly and take appropriate action.

The primary challenge of this project extended beyond hardware development. It involved integrating computer vision models, image-processing workflows, sensing devices, and application software on the MediaTek platform, followed by optimization for real-world product requirements. The goal was to ensure the required recognition accuracy, responsiveness, and long-term system stability.

The overall development process consisted of three main stages:
Model and System Requirements Assessment → AI Inference and Platform Optimization → Hardware Integration and Validation

AI Model Deployment Involves More Than Loading and Running a Model

Running a computer vision model reliably on an embedded device requires careful consideration of the model architecture, inference workflow, computational operators, data precision, memory usage, and hardware acceleration methods. Even if a model loads successfully, its inference speed, recognition accuracy, and system stability may still fall short of actual product requirements.

AI model deployment is therefore not a standalone task. It is a complete system engineering effort encompassing model architecture, inference engines, hardware acceleration, image processing, and embedded software.

Identifying Development Risks Through Model and System Requirements Assessment

Before platform integration began, InnoComm evaluated the model and system requirements by analyzing the model’s network architecture, input and output formats, image pre-processing and post-processing workflows, and computational nodes.

The engineering team determined whether the operators used by the model could be efficiently supported by the MediaTek platform’s inference engine and AI acceleration unit. It also identified issues that could affect inference performance, recognition accuracy, and the development schedule, including:

Dependencies on specific frameworks or libraries
Computational operators with inefficient execution
Numerical precision settings or model structures unsuitable for the target hardware
Excessive memory usage or bandwidth requirements
Additional latency introduced by image pre-processing and post-processing

The purpose of this assessment was not merely to confirm that the model could run. More importantly, it helped the team anticipate potential performance and accuracy differences and establish a clear direction for subsequent optimization.

AI Inference and Platform Optimization

After completing the requirements analysis, InnoComm proceeded with inference workflow implementation and platform optimization. For operators that the platform could not support efficiently, the engineering team replaced, decomposed, or redesigned them according to the model’s characteristics. This enabled workloads to be allocated effectively across the MediaTek platform’s AI acceleration resources.

The work included:

AI inference workflow implementation and integration
Computational operator analysis, adjustment, and optimization
Image pre-processing and inference post-processing adjustments
Model quantization and numerical precision configuration
Memory footprint and data-transfer efficiency optimization
Inference speed, recognition accuracy, and system stability tuning

These adjustments affect one another. For example, more aggressive model quantization can reduce computational load, memory requirements, and power consumption, but may affect recognition accuracy. Increasing the input image resolution can preserve more visual features, but it also increases inference latency and overall system load.

InnoComm therefore helped the customer strike the right balance among recognition accuracy, responsiveness, power consumption, hardware cost, and system resources. The objective was not simply to run the AI functions on the platform, but to ensure they met real-world deployment requirements.

Validating Real-World System Performance on Hardware

After AI inference and platform optimization were completed, the solution still had to be deployed and validated on the actual MediaTek hardware platform.

InnoComm tested inference performance, recognition results, memory usage, system power consumption, and long-term operational stability under different use scenarios on the target device and within the complete hardware and software environment. The team also verified integration among the model, cameras, sensors, network communications, and application software.

Key areas of hardware validation included:

Actual inference speed and event response time
Recognition accuracy across different care scenarios
Power consumption and system temperature during continuous, long-term inference
Stability when multiple sensing devices are connected simultaneously
Recovery capabilities following network interruptions or service failures

Successfully running a model in a development environment is only the first step in building an AI function. The true key to productization is ensuring that it can operate continuously, responsively, and reliably on actual hardware.

Supporting Real-Time Risk Detection and Privacy Protection

In smart care environments, AI systems must identify high-risk behaviors that may lead to falls or other accidents. When an abnormal event occurs, the system must provide real-time alerts and the necessary visual context so caregivers can assess the situation quickly.

Such systems can support fall-risk management, bed and chair status alerts, and privacy-protected live video. Through privacy-processing mechanisms such as image blurring, the system can retain the information needed for event assessment while reducing the risk of directly exposing residents’ images.

The system can use the MediaTek platform’s heterogeneous computing resources to perform image processing and AI inference, then transmit the recognition results to a care management platform or mobile application.

Instead of receiving only a generic notification, caregivers receive alert information that includes the event type and contextual details. This helps them distinguish genuine high-risk events from routine daily activities more quickly and reduces unnecessary alarm fatigue.

Extending AI Inference into Full-System Optimization

The real-world performance of an AI model does not depend solely on the model itself. Every stage—from camera capture, image decoding, pre-processing, and model inference to event determination, privacy processing, and network transmission—can affect the system’s overall response time.

InnoComm therefore combined AI model optimization with embedded system development. The team reviewed the complete data flow from image input to alert output, reduced unnecessary memory copies and data-format conversions, and improved coordination among different computing units.

Smart care devices are generally expected to operate continuously for extended periods. The system must therefore account for the power consumption and heat generated by continuous inference, the connection stability of cameras and sensors, recovery from network disruptions, and memory usage and system reliability over prolonged operation.

Successfully running a model in a laboratory environment is only the starting point. The ability to operate reliably over extended periods in real care environments is what turns an AI proof of concept into a commercial product.

Combining Hardware and Software Expertise to Accelerate Smart Care Product Deployment

This project demonstrates how InnoComm applies AI inference optimization and hardware-software integration expertise to help customers build smart devices that address real-world care needs.

InnoComm provides more than a MediaTek hardware platform. Its end-to-end engineering capabilities encompass AI model optimization, embedded software integration, image processing, hardware design, and comprehensive system validation.

Enabling AI models to run reliably on target hardware—and achieving product-level optimization across accuracy, performance, power consumption, cost, and reliability—is the core value of InnoComm’s Edge AI system integration expertise.