PaperlayerView exact revision on arXiv

Source-linked AI summary

A new design of a fall detection system integrating landmark identification and deep learning techniques

Tri Nhut Do, Thi Thuy Le

arXiv:2610.05749v1cs.CV

TL;DR

Falls are a serious concern for older adults, and existing wearable and ambient systems have limitations. This study combines MediaPipe-based landmark identification with deep learning for fall detection, reporting over 95% accuracy across diverse datasets.

  • Problem

    Existing wearable and ambient fall-detection systems have limitations including discomfort, battery constraints, false positives, and accuracy issues in changing indoor environments.

    Source: §1 · p. 1

  • Method

    The system combines MediaPipe pose estimation and landmark identification with deep-learning prediction to distinguish falls from routine movements.

    Source: §1 · pp. 1–2

  • Results

    Over 95% accuracy across diverse datasets, surpassing traditional wearable systems (80–90%) and ambient systems (85–90%).

    Source: §5 · p. 9

  • Takeaways & Limitations

    The authors present the non-invasive camera-based design as suitable for long-term smart-home monitoring and enhanced user acceptance.

    Source: §5 · p. 9

    Low light reduces keypoint detection accuracy by approximately 10%, while occlusions and single-camera coverage can also result in missed falls.

    Source: §5 · p. 9

Abstract

from arXiv · show

This article introduces an innovative system that integrates landmark identification with deep learning to enhance fall detection accuracy and reliability. By utilizing advanced computer vision techniques, such as Media Pipe for spatial recognition, the system effectively differentiates between routine movements and actual falls. The integration of landmarks with a deep learning prediction algorithm minimizes false alarms, ensuring timely responses to genuine falls. Comprehensive experimentation underscores the system's versatility across various scenarios, emphasizing its potential to improve safety and independence for older adults. The training process demonstrates a steady increase in accuracy, stabilizing by the 40th cycle, while error rates decline significantly during the initial cycles. Real-time experiments, involving both male and female participants aged 8 to 50, recorded a remarkable 95% detection rate of falls, demonstrating the system's effectiveness and promising future applications in elder care and smart health monitoring environments.

1. Introduction

Falls pose a serious public-health concern for older adults, while current fall-detection systems face practical and accuracy limitations. The study proposes a vision-based system combining landmark identification with deep learning for elder care.

1. Introduction

  • Falls are a major public-health concern: the global ageing population is projected to reach 1.5 billion by 2050, and falls claim 646,000 lives annually, predominantly among older adults.In Vietnam, 1.5–1.9 million older adults experience falls annually, and 17% of monthly geriatric admissions at one Ho Chi Minh City medical centre are fall-related.Source: §1 · p. 1
  • Wearables face discomfort, 12–24-hour battery life, and up to 20% false positives, while ambient systems reach 15% false positives in variable indoor settings.The cited systems also struggle to distinguish falls from routine activities or to maintain accuracy in dynamic environments.Source: §1 · p. 1
  • Controlled-setting studies reported 85% accuracy for image-recognition fall detection and 90% precision for CNN classification, but noted real-world or untrained-scenario weaknesses.The introduction identifies a need for contextual awareness, real-time processing, and generalisation to diverse environments.Source: §1 · p. 1
  • The study proposes a vision-based fall-detection system combining landmark identification with deep learning to distinguish falls from routine movements.Its stated objectives also include contextual analysis of spatial keypoints, detection latency below 500 ms, and scalability for IoT-integrated smart homes.Source: §1 · pp. 1–2
  • The system is intended to be non-invasive and user-friendly, with real-time performance and compatibility with IoT ecosystems for elder-care monitoring.The authors describe the design as adaptable to diverse indoor settings and intended to support safety and independence for older adults.Source: §1 · p. 2

2. Background review

The background reviews fall risks and existing detection approaches, then describes MediaPipe pose estimation as a camera-based way to track movement. It also notes environmental constraints and potential directions for extending such systems.

2. Background review

  • Falls can cause fractures and head trauma that reduce older adults’ quality of life and independence; falls account for about 646,000 deaths annually, with older adults most affected.The review frames fall detection as a public-health concern as the population of older adults grows.Source: §2 · p. 2
  • Wearable devices can detect falls but may cause discomfort, have limited battery life, and mistake routine movements for falls; ambient sensors can be affected by lighting, obstacles, and false positives.The review describes both categories as less suitable when comfort or changing environmental conditions are concerns.Source: §2 · p. 2
  • MediaPipe’s Pose Estimation model tracks 33 body keypoints in real time, and prior work describes combining pose tracking with LSTM networks to model movement over time.The review says this temporal modelling can enhance fall-prediction accuracy.Source: §2 · p. 2
  • Camera-based MediaPipe systems use standard cameras and avoid wearable devices, which the review presents as cost-effective and more acceptable for older adults.The review also notes that computer-vision systems can offer a wider field of view than sensor-based systems.Source: §2 · p. 3
  • Poor lighting, camera positioning, or obstructions such as furniture and other people can cause missed detections or false alarms in MediaPipe-based systems.The review says calibration and advanced image processing may mitigate these limitations.Source: §2 · p. 3
  • Suggested future directions include edge-based IoT processing and multi-camera views, as well as GAN-generated training data for broader fall scenarios.The review presents these as potential ways to support real-time alerts, wider environmental coverage, and adaptation to varied conditions.Source: §2 · p. 3
  • The review presents MediaPipe pose estimation with deep learning as a non-invasive, camera-based approach for distinguishing routine movement from falls in elder care.It identifies lighting and obstructions as continuing challenges for this approach.Source: §2 · p. 3

3. System design

The system processes camera video through fall detection, alerts, and an alarm, using MediaPipe pose keypoints and an LSTM to analyze movement sequences. Its training data comprises labeled fall and no-fall sequences, with a 70/30 training-test split.

3. System design

  • Video passes through fall-detection processing and an alert module before a confirmed fall triggers a sound alarm.The processing analyzes motion patterns such as changes in posture, velocity, or position.Source: §3 · pp. 3–4

3.1. Fall Detection Processing design

  • The proposed pipeline extracts 33 pose keypoints from 1080p RGB video at 30 frames per second, then normalizes, filters, and segments them into one-second windows for LSTM analysis.The authors claim their preprocessing and adaptive modeling generalize across diverse indoor environments, contrasting with earlier controlled-setting or laboratory results.Source: §3.1 · p. 4, Figure 3 · p. 4

3.2. MediaPipe Pose

  • MediaPipe Pose uses CNNs to locate 33 body keypoints, which provide real-time posture and movement information.The keypoints include positions such as the head, shoulders, hips, knees, and ankles.Source: §3.2 · pp. 4–5

3.3. Modelling - Data capturing and labelling

  • The dataset combines self-collected participants with Le2i and Multiple Cameras Fall Dataset coordinates, representing each of 520 videos as 50 frames of 132 features.Each 50-frame segment is divided into overlapping 30-frame sequences shifted by ten frames; the sequences are labeled fall or no fall.Source: §3.3 · p. 5

3.4. Modelling - Training and testing dataset

  • Training uses 26,000 feature arrays, with 70% assigned to training and 30% to testing; evaluation includes accuracy, F1 score, precision, recall, and MCC.The stated inputs represent 33 keypoints by four values: x, y, z, and display confidence.Source: §3.4 · p. 6

3.5. Modelling - Long Short-term Memory

  • LSTM processes sequences through cell states and forget, input, and output gates, supporting retention of long-term information.The forget gate controls what to retain from the previous cell state, the input gate controls what new information is added, and the output gate selects information for the current output.Source: §3.5 · p. 6

4. Experimental results and evaluation

Real-time tests detected 19 of 20 falls, while validation accuracy exceeded 99%. The experiments also report training convergence patterns and the system setup used for evaluation.

4. Experimental results and evaluation

  • The evaluation section describes experimental scenarios involving computer vision, deep learning, and real-world testing.The stated aims include dataset development, model evaluation, and testing different fall types.Source: §4 · p. 6

4.1. System initial setup

  • Experiments used a Windows 11 computer, a camera operating at 25 fps and 2 MP resolution, and a TensorFlow/Keras LSTM model.The computer had an 11th-generation Intel Core i5-1135G7 processor, 16 GB RAM, and Intel Iris Xe Graphics.Source: §4.1 · p. 7

4.2. Training validation and experimental results evaluation

  • The test confusion matrix recorded 116 true positives, 0 false positives, 1 false negative, and 111 true negatives for fall and no-fall classifications.Fig. 8 presents actual versus predicted classifications; specific matrix values are listed in Table 2.Source: §4.2 · pp. 7–8, Figure 8 · p. 7, Table 2 · p. 7
  • Training and testing accuracy rose steadily and began stabilizing by cycle 40, while error rates fell sharply in the first ten cycles and stabilized around cycle 45 onward.Accuracy and error rates are shown in Figs. 9 and 10, respectively.Source: §4.2 · p. 8, Figure 9 · p. 8, Figure 10 · p. 8
  • Over 99% accuracy was achieved on the test dataset after 80 iteration cycles.The model parameters obtained after 80 cycles are listed in Table 3.Source: §4.2 · p. 8, Table 3 · p. 8
  • 95% of 20 real-time fall actions were detected among participants aged 8–50, with processing times of 0.02–0.11 seconds and response times under 0.09 seconds.Each 2–5 minute experiment involved one participant performing normal walking and at least one unexpected fall action.Source: §4.2 · p. 8

5. Conclusions

The system reports strong fall-detection accuracy across diverse datasets and varied environments, while identifying environmental, coverage, computational, and privacy constraints. Proposed future work targets robustness, efficiency, response time, and data protection.

5. Conclusions

  • Over 95% accuracy across diverse datasets surpasses traditional wearable systems (80-90%) and ambient systems (85-90%).The system uses MediaPipe pose estimation and LSTM networks to model 33 body keypoints and distinguish falls from daily activities, with a false-positive rate below 5% versus 15-20% for sensor-based methods.Source: §5 · p. 9
  • Across care facilities and lighting conditions of 50-1000 lux, the approach yields F1-scores above 0.93, outperforming lab-constrained vision systems with F1-scores of 0.85-0.90.Its non-invasive, camera-based design enhances user acceptance, making it ideal for long-term smart home monitoring.Source: §5 · p. 9
  • Low light, overexposure, occlusion, single-camera coverage, LSTM computation, and continuous-video privacy constrain performance or deployment.Reported examples include a roughly 10% keypoint-accuracy drop below 50 lux, 5-10% of falls missed outside the field of view in areas over 20 m², and inference delays on low-power devices.Source: §5 · p. 9
  • Future work proposes infrared or depth cameras, multi-camera coverage, lightweight models, edge computing, and privacy-preserving processing.The passage projects over 97% low-light accuracy with infrared or depth cameras, approximately 30% faster inference with lightweight models, and sub-one-second caregiver alerts using IoT-enabled edge computing.Source: §5 · p. 9

CRediT author statement

Both authors contributed to conceptualization, methodology, and reviewing and editing; their other listed responsibilities differ.

CRediT author statement

Loading 2610.05749v1…