The global demographic shift toward an aging population presents substantial challenges for healthcare systems worldwide. Elderly individuals often face mobility issues, cognitive decline and social isolation, necessitating the development of assistive technologies that can support independent living. Intelligent robots equipped with Computer Vision (CV) and Machine Learning (ML) technologies offer a promising solution to these challenges, providing not only physical assistance but also emotional support. CV enables robots to perceive and understand visual data, while ML allows them to analyze large datasets and learn from experiences. This facilitates tasks such as fall detection, activity recognition and personalized health assessments [1].  

While existing systems offer basic monitoring capabilities, a significant gap remains in integrating real-time multi-modal data with continuous, adaptive learning to achieve highly accurate, personalized and proactive care.

This research focuses on the development of a robotic system that tightly integrates CV and ML for elderly care. The system monitors daily activities, detects falls and provides companionship, ultimately improving the quality of life for elderly individuals. By continuously adapting to specific needs, this system provides a tailored, proactive approach to elderly care, substantially reducing the burden on human caregivers [2].

• Sawik, [3]. "Robots for Elderly Care: Review, Multi-Criteria Optimization." PMC. This study provides an in-depth review of the current state of robotic systems designed for elderly care, discussing multi-criteria optimization approaches to improve robot performance in various settings

• Gaya-Morey et al. [4]. "Deep Learning for Computer Vision-Based Activity Recognition and Fall Detection of the Elderly: A Systematic Review." Applied Intelligence. This review focuses on the application of deep learning techniques in activity recognition and fall detection for elderly individuals, providing a comprehensive overview of recent advances and challenges

• Feng, G. [1]. "Artificial Intelligence in Chronic Disease Management for the Elderly." PMC.
  Feng’s work explores the role of artificial intelligence in managing chronic diseases in elderly populations, emphasizing AI's potential to improve care and outcomes

• Wong, [2]. "Adoption of Artificial Intelligence–Enabled Robots in Long-Term Care Homes." JMIR Aging. This paper discusses the barriers and challenges in adopting AI-enabled robots in long-term care homes, providing insights into how these technologies can be integrated into elderly care practices

• Giansanti, [5]. "Integrating AI and Assistive Technologies in Healthcare." PMC. Giansanti explores the integration of AI and assistive technologies in healthcare, focusing on their potential to enhance elderly care, particularly in dementia care.

• Ji, [6]. "A Networked Intelligent Elderly Care Model Based on Machine Learning." Research.
  Ji proposes a networked intelligent elderly care model that uses machine learning to track elderly health conditions and adjust care plans accordingly, offering personalized support

• Aldawsari, [7]. "Intervention Scenarios and Robot Capabilities for Support in Elderly Care." Science Direct. This article presents various intervention scenarios where robots assist the elderly, including companionship, monitoring health and providing mobility assistance, with a focus on the capabilities of current robotic systems

Comparison with Our Work

While previous studies such as those by Feng [1] and Giansanti [5] offer significant insights into AI and assistive technology applications for elderly care, they often focus on chronic disease management or general assistive technology rather than real-time, low-latency and multi-modal fall detection. Specifically, our work distinguishes itself by integrating real-time activity recognition, fall detection and personalized health predictions through a combination of computer vision and machine learning algorithms. Unlike systems focusing on single-sensor data, our approach fuses visual (RGB-D) and auditory data for robust recognition and critically, features continuous learning capabilities that improve the robot's performance over time based on user interaction and collected data. This adaptive mechanism addresses the static nature and slower response times observed in many preceding systems.

Proposed Work

System Overview and Components: The proposed system consists of a mobile robot equipped with an RGB-D camera for depth perception and a microphone array for audio input. The robot utilizes computer vision algorithms to process visual data and machine learning models to analyze activity patterns and detect falls. The key technological components include:

• Computer Vision (CV): Utilization of Convolutional Neural Networks (CNNs) for activity recognition and object detection

• Machine Learning (ML): Implementation of Recurrent Neural Networks (RNNs) for sequential data analysis and predictive modeling

• Sensor Integration: Fusion of visual and auditory data to improve accuracy and robustness

• User Interface: Development of a user-friendly interface for caregivers to monitor and interact with the system 

• The system is designed to operate in real-time, providing immediate feedback and alerts to caregivers when necessary

Proposed Algorithm and Work Details

Algorithm Overview: The proposed algorithm integrates computer vision and machine learning techniques to develop a robot capable of assisting in elderly care. The key tasks are fall detection, activity recognition and providing timely alerts. The primary algorithm design consists of two main components:  

• Computer Vision Module: This module uses RGB-D (depth + color) cameras to capture visual data. The data is processed to detect and track human movements in real-time, using object detection and activity recognition techniques

• Machine Learning Module: This module applies supervised learning models, particularly Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), to identify and predict fall events and daily activities. The algorithm is trained on a labeled dataset of elderly activity and fall scenarios

Key Steps in the Proposed Algorithm

• Data Collection: Use RGB-D cameras to collect real-time video and depth data, capturing both color and depth information for improved detection. Activity data is gathered from labeled datasets of elderly individuals performing activities (e.g., walking, sitting, standing and falling)

• Data Preprocessing: Preprocess the video and depth data to extract key features related to human posture, movement and interactions. Specific key features extracted include joint angles, bounding box aspect ratios and velocity vectors. Image segmentation techniques are used to isolate human figures and detect spatial features

• Activity Recognition: Apply CNNs to analyze static and dynamic patterns in video data. Activities such as walking, sitting, standing, eating and moving are identified. RNNs are utilized to model temporal dependencies and sequence data, enabling the recognition of complex activities over time

• Fall Detection: Train the algorithm to detect abnormal movements associated with falls using labeled fall data. Data fusion from both the depth and visual information (such as sudden shifts in position or a person’s rapid fall) is crucial for improving fall detection accuracy and reducing false positives 

• Decision Making and Alerts: Upon detecting a fall or observing unusual behavior, the system sends real-time alerts to caregivers through a mobile application. Based on activity patterns, the system also suggests health-related recommendations and provides preventive measures

• Adaptive Learning: The system continuously learns from new data, refining its fall prediction and activity recognition models based on past events. Reinforcement learning will be employed, utilizing a reward function that penalizes false alarms and rewards accurate, timely intervention, to adjust the robot's interactions with the elderly person based on feedback and preferences

System Architecture

The architecture consists of several key components working together to create a responsive and intelligent robot.  

• Sensors and Data Collection Layer: Includes RGB-D Cameras for real-time activity and fall detection. (Optional Wearable Sensors for heart rate, body posture, etc. are for future integration)

• Preprocessing Layer: Data Processing Unit handles noise removal and feature extraction. The Computer Vision Module processes video frames to track movements and activities

• Machine Learning Layer: Model Training Unit trains CNNs and RNNs. The Prediction Unit deploys the trained models for real-time decision-making

• Action and Communication Layer: Alert System sends notifications to caregivers. User Interface allows caregivers to monitor the robot's performance

• Data Flow: Data is captured by the RGB-D Cameras and immediately routed to the Preprocessing Layer for feature extraction. The processed features are then fed into the Machine Learning Layer for real-time activity recognition and fall detection. The resultant prediction triggers the Action and Communication Layer, sending alerts or providing recommendations via the Alert System

This Figure 1, summarizes the main steps from data collection to alert generation, highlighting how the system processes input and generates real-time responses.

Figure 1: Flowchart of Proposed Algorithm

Environment, using synthetic data to evaluate key performance metrics such as fall detection, activity recognition, real-time processing time and improvement in fall detection accuracy.

Fall Detection Accuracy

The algorithm achieved a 95% accuracy rate in detecting simulated falls. The false-positive rate, where normal activities were mistakenly classified as falls, was acceptably low at 5%.

• Green Bar: Represents the percentage of falls correctly detected by the system

• Red Bar: Represents the false positives or normal activities mistakenly classified as falls

Activity Recognition

• The system operated with a processing time of 0.2 seconds (200 milliseconds) per frame. This fast-processing time is critical for real-time applications in elderly care, ensuring that life-saving alerts are sent with minimal delay in a critical fall scenario

• Blue Bar: Represents correct activity recognition

• Orange Bar: Represents incorrect recognition, showing where the system failed to recognize the activity accurately

Improvement in Fall Detection Accuracy

• The adaptive learning feature resulted in a 15% improvement in fall detection accuracy after 30 days of continuous learning and data collection. The system was able to adjust its sensitivity based on the individual's unique activity patterns

• The purple bar represents the time taken for processing a single frame (200 milliseconds), which is considered fast for real-time applications (Figure 2)

Preliminary testing of the system in a controlled environment yielded the following results:

• Fall Detection: The system achieved an accuracy rate of 95% in detecting simulated falls

• Activity Recognition: The robot correctly identified daily activities with an accuracy of 92%

• User Interaction: The interface received positive feedback from caregivers regarding usability and functionality

These results demonstrate the feasibility of integrating computer vision and machine learning in robotic systems for elderly care.

The following simulation results were obtained to evaluate the robot's performance in a controlled environment using synthetic data.

Fall Detection

• The system achieved 95% accuracy in detecting falls. This result was obtained by testing the algorithm on 1,000 fall scenarios simulated in a home environment

• The false positive rate was found to be low (5%), as the system correctly identified normal movements and did not misclassify them as falls

Activity Recognition

• The activity recognition system was trained on 5,000 labeled data points representing various activities such as walking, sitting, eating and standing

• The system achieved an overall accuracy of 92% in recognizing these activities in real-time video streams

• Specific activities, such as eating and standing up from a sitting position, were the most accurately detected with a 98% recognition rate

Real-Time Processing

The system operated in real-time with less than 200 milliseconds processing time per frame, ensuring timely alerts and responsiveness

Adaptive Learning

The algorithm was able to adjust its sensitivity to detect falls based on individual activity patterns. This learning process resulted in a 15% improvement in fall detection accuracy after 30 days of continuous data collection.

Comparison with Prior Works

Table 1, comparing our robot's performance with previous research in terms of fall detection accuracy, activity recognition and system performance.

Our system outperforms all previous benchmark systems in fall detection accuracy, achieving 95%. Furthermore, our ability to process video frames in 200 milliseconds compared to the 500 milliseconds taken by previous systems (like Robot 1) demonstrates superior efficiency for real-time response. The unique adaptive learning feature resulted in a 15% improvement in fall detection accuracy over 30 days, a crucial capability absents in static systems like Robot 3.

Figure 2: Simulation Results

Table 1: Comparing Our Robot's Performance with Previous Research in Terms of Fall Detection Accuracy, Activity Recognition and System Performance

The integration of computer vision and machine learning in robotic systems offers several advantages for elderly care: 

• Autonomy: Robots can operate independently, reducing the need for constant human supervision 

• Personalization: Machine learning models can adapt to individual users' needs and preferences, resulting in more accurate and non-intrusive monitoring

• Efficiency: Real-time processing allows for immediate responses to emergencies

However, challenges remain, including:  

• Privacy: Ensuring the security of personal data collected by the system. We will mitigate this by implementing on-device processing where possible, anonymizing activity data and utilizing secure, encrypted cloud transmission for long-term health pattern analysis

• Acceptance: Gaining the trust and acceptance of elderly individuals and caregivers

• Integration: Seamlessly incorporating the system into existing healthcare infrastructures

The high accuracy and low latency demonstrated by our system suggest a profound potential for reducing the emotional and physical burden on human caregivers by significantly decreasing false alarms and guaranteeing timely intervention. Addressing the aforementioned challenges is crucial for the successful deployment of intelligent robots in elderly care settings

The proposed algorithm successfully integrates computer vision and machine learning techniques to develop a robot capable of assisting in elderly care. Simulation results demonstrate the algorithm's high efficiency in fall detection (95% accuracy) and activity recognition (92% accuracy). This work provides a significant advancement by incorporating a continuous adaptive learning mechanism, which improved fall detection accuracy by 15%, demonstrating the system's ability to provide truly personalized and proactive care. Future work will focus on refining the system's capabilities, testing in real-world environments and improving interaction models based on continuous learning from the elderly person’s data.

1. Feng, G. “Artificial intelligence in chronic disease management for the elderly.” PMC, 2025.

2. Wong, K.L.Y. “Adoption of artificial intelligence–enabled robots in long-term care homes.” JMIR Aging, 2024.

3. Sawik, B. “Robots for elderly care: Review, multi-criteria optimization.” PMC, 2023.

4. Gaya-Morey, F. et al. “Deep learning for computer vision based activity recognition and fall detection of the elderly: A systematic review.” Applied Intelligence, 2024.

5. Giansanti, D. “Integrating AI and assistive technologies in healthcare.” PMC, 2025.

6. Ji, W. “A Networked Intelligent Elderly Care Model Based on Machine Learning.” Research, 2025.

7. Aldawsari, S. “Intervention scenarios and robot capabilities for support in elderly care.” Science Direct, 2024.