[Paper Review] Radar Classification of Contiguous Activities of Daily Living

Ronny G. Guendel|arXiv (Cornell University)|Dec 17, 2019

Advanced SAR Imaging Techniques42 references4 citations

TL;DR

This paper proposes a radar-based classification system for contiguous Activities of Daily Living (ADLs), using the Radon transform on range-maps to detect translation motion and an energy-based Power Burst Curve (PBC) to identify in-place motion onsets and offsets. By leveraging the human ethogram to dynamically reduce classification classes based on current motion states, the method achieves up to 99.7% classification accuracy in forward time and 98.9% in backward time, outperforming global classifiers on complex motion sequences like walking-falling-sitting.

ABSTRACT

We consider radar classifications of Activities of Daily Living (ADL) which can prove beneficial in fall detection, analysis of daily routines, and discerning physical and cognitive human conditions. We focus on contiguous motion classifications which follow and commensurate with the human ethogram of possible motion sequences. Contiguous motions can be closely connected with no clear time gap separations. In the proposed motion classification approach, we utilize the Radon transform applied to the radar range-map to detect the translation motion, whereas an energy detector is used to provide the onset and offset times of in-place motions, such as sitting down and standing up. It is shown that motion classifications give different results when performed forward and backward in time. The number of classes, thereby classification rates, considered by a classifier, is made variable depending on the current motion state and the possible transitioning activities in and out of the state. Motion examples are provided to delineate the performance of the proposed approach under typical sequences of human motions.

Motivation & Objective

• Address the challenge of classifying contiguous, non-separable ADLs such as walking followed by falling or sitting, which lack clear temporal boundaries.
• Overcome limitations of traditional radar classifiers that treat all ADLs uniformly, regardless of motion context or sequence.
• Improve classification accuracy for ambiguous motion transitions—especially walking-falling and walking-sitting—by incorporating temporal context and motion state dynamics.
• Enable robust classification in both forward and backward time directions to enhance reliability and detect subtle motion transitions.
• Reduce computational load and improve accuracy by dynamically adjusting the number of classes based on the current motion state and possible transitions, using the human ethogram as a guide.

Proposed method

• Apply the Radon transform to radar range-maps to detect translation motion, such as walking, by identifying linear energy patterns corresponding to movement.
• Use a Power Burst Curve (PBC) energy detector to detect onset and offset times of in-place motions like sitting down, standing up, or falling.
• Model human motion as a state machine with four primary states: standing, sitting, walking, and laying, and define transition actions (e.g., standing up, falling) between them using the human ethogram.
• Dynamically reduce the number of classification classes at each state by considering only possible incoming and outgoing transitions, improving classifier focus and accuracy.
• Employ 2D Principal Component Analysis (2D-PCA) on range-maps and micro-Doppler spectrograms, followed by Nearest Neighbor (NN) classification, with the number of principal components adjusted per state.
• Perform classification in both forward and backward time directions to evaluate temporal asymmetry and improve confidence in motion recognition.

Experimental results

Research questions

• RQ1 How does classifying ADLs in both forward and backward time directions affect classification accuracy for contiguous motion sequences?
• RQ2 To what extent does reducing the number of active classes based on motion state and ethogram constraints improve classification performance compared to using all ADL classes?
• RQ3 Can the Radon transform and Power Burst Curve (PBC) effectively distinguish translation motion from in-place motions in radar-sensed data?
• RQ4 How do motion sequences involving ambiguous transitions—such as walking into a fall or sitting—impact classification accuracy, and can the proposed method mitigate this?
• RQ5 What is the performance gain of using a state-dependent classifier with dynamic class reduction compared to a fixed, global classifier across all ADLs?

Key findings

• The proposed method achieved a forward-time classification accuracy of 99.70% on Example-3 (picking up an object), outperforming the backward-time method (99.60%) and global classifiers.
• For the walking-falling-sitting sequence (Example-2), the proposed method achieved 96.95% accuracy in forward time and 98.90% in backward time, significantly outperforming the global classifier (92.70% forward, 96.70% backward).
• In the walking-falling sequence (Example-1), the proposed method reached 91.70% accuracy in forward time and 97.20% in backward time, demonstrating that backward classification often yields higher accuracy for ambiguous transitions.
• The average classification rate across all examples was 98.26% (forward) and 98.60% (backward) when using the global classifier, but the proposed method achieved up to 99.70% in forward time on specific sequences.
• The use of a state-based classifier with dynamic class reduction based on the human ethogram significantly improved performance on complex, contiguous motion sequences compared to a fixed, all-class classifier.
• The method demonstrated that motion classification results differ substantially when performed in forward versus backward time, indicating that bidirectional classification enhances reliability and confidence in activity recognition.