Towards Explainable, Privacy-Preserved Human-Motion Affect Recognition

TL;DR

This paper presents a novel deep learning approach for emotion recognition from human gait that preserves subject privacy by disentangling biometric features, outperforming traditional methods and providing explainability through Grad-CAM.

Contribution

The work introduces a multi-encoder autoencoder with transfer learning to disentangle emotions from biometrics in gait data, enhancing privacy and recognition accuracy.

Findings

Up to 7% improvement in emotion recognition accuracy.

Effective disentanglement of biometric features from gait data.

Grad-CAM provides interpretable insights into model decisions.

Abstract

Human motion characteristics are used to monitor the progression of neurological diseases and mood disorders. Since perceptions of emotions are also interleaved with body posture and movements, emotion recognition from human gait can be used to quantitatively monitor mood changes. Many existing solutions often use shallow machine learning models with raw positional data or manually extracted features to achieve this. However, gait is composed of many highly expressive characteristics that can be used to identify human subjects, and most solutions fail to address this, disregarding the subject's privacy. This work introduces a novel deep neural network architecture to disentangle human emotions and biometrics. In particular, we propose a cross-subject transfer learning technique for training a multi-encoder autoencoder deep neural network to learn disentangled latent representations of human motion features. By disentangling subject biometrics from the gait data, we show that the subject's privacy is preserved while the affect recognition performance outperforms traditional methods. Furthermore, we exploit Guided Grad-CAM to provide global explanations of the model's decision across gait cycles. We evaluate the effectiveness of our method to existing methods at recognizing emotions using both 3D temporal joint signals and manually extracted features. We also show that this data can easily be exploited to expose a subject's identity. Our method shows up to 7% improvement and highlights the joints with the most significant influence across the average gait cycle.