[Paper Review] Identifying Emotion from Natural Walking

Liqing Cui, Shun Li+3 more|arXiv (Cornell University)|Aug 3, 2015

Anomaly Detection Techniques and Applications8 references3 citations

TL;DR

This paper proposes a method to identify human emotions from natural walking using accelerometer data from smartphones worn on the wrist and ankle. By applying moving average filtering, extracting 114 time- and frequency-domain features, and using PCA for dimensionality reduction, the study achieves up to 90.31% accuracy in classifying anger and 89.76% for happiness using SVM, with ankle data outperforming wrist data, demonstrating that gait patterns reliably reflect affective states.

ABSTRACT

Emotion identification from gait aims to automatically determine persons affective state, it has attracted a great deal of interests and offered immense potential value in action tendency, health care, psychological detection and human-computer(robot) interaction.In this paper, we propose a new method of identifying emotion from natural walking, and analyze the relevance between the traits of walking and affective states. After obtaining the pure acceleration data of wrist and ankle, we set a moving average filter window with different sizes w, then extract 114 features including time-domain, frequency-domain, power and distribution features from each data slice, and run principal component analysis (PCA) to reduce dimension. In experiments, we train SVM, Decision Tree, multilayerperception, Random Tree and Random Forest classification models, and compare the classification accuracy on data of wrist and ankle with respect to different w. The performance of emotion identification on acceleration data of ankle is better than wrist.Comparing different classification models' results, SVM has best accuracy of identifying anger and happy could achieve 90:31% and 89:76% respectively, and identification ratio of anger-happy is 87:10%.The anger-neutral-happy classification reaches 85%-78%-78%.The results show that it is capable of identifying personal emotional states through the gait of walking.

Motivation & Objective

• To investigate whether emotional states can be reliably inferred from natural walking using wearable accelerometer data.
• To compare the effectiveness of wrist versus ankle sensor placement in emotion identification.
• To evaluate the impact of data preprocessing, particularly moving average filtering with varying window sizes, on classification accuracy.
• To determine which machine learning models perform best in classifying emotions from gait features.
• To explore the feasibility of using everyday smartphone sensors for continuous, non-invasive affective state monitoring.

Proposed method

• Accelerometer data was collected from participants walking naturally while wearing smartphones on the wrist and ankle.
• A moving average filter with variable window size w (3 or 5) was applied to smooth the raw acceleration signals and reduce noise.
• A total of 114 features—time-domain, frequency-domain, power, and distribution features—were extracted from each filtered data segment.
• Principal component analysis (PCA) was used to reduce dimensionality and eliminate redundant information from the high-dimensional feature set.
• Five classification models—SVM, Decision Tree, Multilayer Perceptron, Random Tree, and Random Forest—were trained and evaluated using 10-fold cross-validation.
• Performance was compared across different window sizes (w=3, w=5) and sensor locations (wrist vs. ankle) to assess impact on emotion classification accuracy.

Experimental results

Research questions

• RQ1 Can emotional states such as anger, happiness, and neutrality be accurately classified from natural walking using smartphone accelerometer data?
• RQ2 Does the placement of the sensor (wrist vs. ankle) significantly affect the accuracy of emotion classification?
• RQ3 How does the size of the moving average filter window (w) influence the performance of emotion classification models?
• RQ4 Which machine learning model yields the highest classification accuracy for affective state recognition from gait features?
• RQ5 To what extent does data preprocessing, including noise reduction and feature extraction, impact the final classification performance?

Key findings

• The SVM classifier achieved the highest accuracy, reaching 90.31% for identifying anger and 89.76% for identifying happiness when using ankle accelerometer data.
• Ankle data consistently outperformed wrist data in emotion classification, with the best overall accuracy of 90.31% achieved at a moving average window size of w=5.
• The confusion matrix results indicate that happiness is most frequently misclassified as anger, and vice versa, suggesting overlapping gait patterns between these two emotions.
• Neutral affect was the most accurately classified emotion, with an accuracy of 86% for wrist and 86% for ankle data when w=3, and 78% for ankle at w=5.
• The anger-neutral-happy classification achieved 85% for anger, 78% for neutral, and 78% for happiness when using ankle data with w=5, indicating moderate but consistent performance across all three classes.
• The study confirms that gait patterns contain discriminative features for affective states, and that preprocessing steps such as moving average filtering and PCA significantly enhance model performance.