[Paper Review] Scale-Invariant Local Descriptor for Event Recognition in 1D Sensor Signals

Jierui Xie, Mandis Beigi|arXiv (Cornell University)|May 28, 2011

Time Series Analysis and Forecasting4 references4 citations

TL;DR

This paper proposes a scale-invariant local descriptor for 1D sensor signals that enables robust event recognition across varying temporal scales, such as fast and slow walking. By leveraging scale-space theory and a high-level shape encoding scheme, the method extracts invariant features and employs an R-metric-based classifier (SIC-R) that combines keypoint matches and Dynamic Time Warping scores, achieving over 90% accuracy on diverse 1D sensor data from PIR, accelerometer, and seismic sensors.

ABSTRACT

In this paper, we introduce a shape-based, time-scale invariant feature descriptor for 1-D sensor signals. The time-scale invariance of the feature allows us to use feature from one training event to describe events of the same semantic class which may take place over varying time scales such as walking slow and walking fast. Therefore it requires less training set. The descriptor takes advantage of the invariant location detection in the scale space theory and employs a high level shape encoding scheme to capture invariant local features of events. Based on this descriptor, a scale-invariant classifier with "R" metric (SIC-R) is designed to recognize multi-scale events of human activities. The R metric combines the number of matches of keypoint in scale space with the Dynamic Time Warping score. SICR is tested on various types of 1-D sensors data from passive infrared, accelerometer and seismic sensors with more than 90% classification accuracy.

Motivation & Objective

• To address the challenge of recognizing human activities in 1D sensor signals that vary significantly in temporal duration, such as walking fast vs. slow.
• To reduce the need for extensive training data by enabling generalization across different time scales.
• To develop a feature descriptor that is invariant to temporal scaling while preserving discriminative local shape characteristics.
• To design a classifier capable of matching events across scale variations using a hybrid metric combining keypoint correspondence and temporal alignment.

Proposed method

• The method uses scale-space theory to detect stable keypoint locations across multiple temporal scales in 1D signals.
• A high-level shape encoding scheme captures local signal morphology around detected keypoints, ensuring invariance to scale changes.
• The SIC-R classifier computes an R metric that combines the number of matched keypoints in scale space with a Dynamic Time Warping (DTW) score for temporal alignment.
• The R metric is used to compare query events against training events, enabling robust classification under scale variation.
• The descriptor is evaluated on 1D signals from passive infrared, accelerometer, and seismic sensors to test cross-sensor robustness.

Experimental results

Research questions

• RQ1 Can a local feature descriptor be designed to be invariant to temporal scaling in 1D sensor signals?
• RQ2 How can keypoint-based features be encoded to preserve shape information across scale variations?
• RQ3 Can a hybrid metric combining keypoint matching and DTW improve classification accuracy on multi-scale events?
• RQ4 To what extent does the proposed method generalize across different types of 1D sensor data?

Key findings

• The proposed scale-invariant local descriptor achieves over 90% classification accuracy on diverse 1D sensor data, including passive infrared, accelerometer, and seismic signals.
• The SIC-R classifier outperforms baseline methods by effectively combining keypoint matches and DTW alignment, enhancing robustness to scale variations.
• The method reduces the need for large, multi-scale training sets by generalizing across different temporal durations of the same activity.
• The high-level shape encoding scheme successfully captures discriminative local features while maintaining invariance to scale changes.