Using Temporal Covariance of Motion and Geometric Features via Boosting for Human Fall Detection

Syed Farooq Ali, Reamsha Khan, Arif Mahmood, Malik Tahir Hassan, Moongu Jeon

TL;DR

This paper introduces a novel, real-time human fall detection system that leverages temporal covariance of motion and geometric features. By combining these features with a boosted J48 classifier, the system achieves high accuracy and speed, outperforming existing methods on standard datasets. This advancement is crucial for developing more reliable automated systems to assist vulnerable individuals.

What Problem Does This Paper Address?

The increasing elderly population worldwide presents a growing challenge: the risk of falls and the subsequent need for timely medical intervention. While existing fall detection systems exist, they often suffer from inaccuracies, unreliability, or limitations like dependence on wearable devices that can be forgotten or become ineffective. This paper addresses the need for a robust, accurate, and fast vision-based fall detection system that can operate in real-time using standard surveillance cameras, thereby improving the safety and well-being of individuals, particularly the elderly.

Methods

The proposed system employs a multi-stage approach:

  1. Foreground Detection: Standard background subtraction techniques are used to isolate moving objects (potential persons) from the background.
  2. Feature Extraction: The core innovation lies in extracting a set of temporal and spatial variance-based features. These include:
  3. Temporal Variance of Aspect Ratio: Captures changes in the bounding box's width-to-height ratio, which is unique during a fall.
  4. Temporal Variance of Person Angle: Measures the change in the person's orientation (angle of the major axis of an ellipse fitted to the blob) over time.
  5. Temporal Variance of Motion Vector: Tracks the variation in the body's speed, which increases significantly during a fall.
  6. Temporal Variation of Shape Deformations: Utilizes the upper half area of the bounding box to detect changes in body orientation.
  7. Temporal Variation in Geometric Center Position: Monitors the movement of the object's centroid.
  8. Temporal Variation of the Ellipse Ratio: Analyzes the ratio of the major to minor axes of an ellipse, serving as a scale-invariant feature.
  9. Classification: An ensemble learning strategy, specifically boosting, is applied. The authors chose to boost the J48 decision tree classifier (an extension of ID3) due to its efficiency, robustness, and ability to handle various data types. They also experimented with AdaBoost as a comparison. The boosting approach sequentially trains classifiers, giving more weight to misclassified samples, leading to improved overall performance.

Datasets

The proposed system was evaluated on two publicly available datasets:

  • Multiple Cameras Fall (MCF) Dataset: This dataset contains 20 videos captured from different camera positions, featuring various types of falls.
  • UR Fall Detection (URFD) Dataset: This dataset includes frontal and overhead video sequences from Kinect sensors, featuring falls from standing and sitting positions.

The paper does not introduce a new dataset but uses these existing benchmarks to compare its performance against state-of-the-art methods.

Key Insights and Ideas

  • Temporal Covariance is Key: The paper strongly emphasizes that temporal variations in motion and geometric features, rather than static snapshots, are crucial for accurately detecting falls. This captures the dynamic nature of falling.
  • Feature Robustness to Viewpoint: A significant finding is that the proposed features remain relatively consistent and exhibit similar temporal variations across different camera views. This is a critical advantage for real-world surveillance systems where camera angles can vary.
  • Boosting J48 for Efficiency and Accuracy: The choice of a boosted J48 classifier is highlighted as a strategic decision. It balances the need for high accuracy with computational efficiency, making it suitable for real-time applications, especially when compared to more complex deep learning models that require extensive training data and computational resources.
  • Distinguishing Falls from Similar Actions: The experiments, particularly Experiment 3 with a three-class classification (fall, sitting, no fall), demonstrate the system's ability to differentiate between rapid falls and slower actions like sitting down, a common challenge for fall detection systems.

Key Results

The proposed system, particularly the PA_B-J48 (Boosted J48) approach, demonstrated superior performance on both datasets:

  • MCF Dataset: Achieved percentage accuracies of 99.2% (for 2 classes) and 99.25% (for 3 classes). In Experiment 1, it outperformed existing methods with accuracies up to 99.86%.
  • URFD Dataset: Showed high accuracy, with 99.13% on the frontal sequence and 99.03% on the overhead sequence.
  • Speed: The proposed algorithm offers a good balance of speed and accuracy. While PA_Ada was faster, PA_B-J48 was still competitive and significantly more accurate than many existing methods. For instance, on the MCF dataset (Experiment 1), PA_B-J48 was faster than Osu and Fo.

Limitations and Future Work

  • Challenging Datasets: The authors acknowledge that the proposed system's performance might degrade on more challenging datasets that involve multiple people, occlusions, similar clothing colors to the background, and non-Lambertian lighting.
  • Future Work:
    • Improving Performance and Time Efficiency: Further research is suggested to enhance the system's capabilities, especially for complex scenarios.
  • Novel Features and Video Repositories: Developing more novel and robust features and creating larger video repositories are identified as key areas for improvement.
  • Night Vision Functionality: Integrating night vision capabilities is proposed as an important feature for outdoor fall detection systems.
  • Parallel Implementation: For larger datasets, a parallel implementation of the PA_B-J48 algorithm is suggested to significantly reduce execution time.

Summary updated

BibTeX
@inproceedings{ali2018using,
  author = {Syed Farooq Ali and Reamsha Khan and Arif Mahmood and Malik Tahir Hassan and Moongu Jeon},
  title = {{Using Temporal Covariance of Motion and Geometric Features via Boosting for Human Fall Detection}},
  booktitle = {Sensors},
  year = {2018},
  url = {https://openreview.net/forum?id=ejnQduQbvlG},
}
Figure