Gait Recognition via Disentangled Pose-Appearance Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing gait recognition methods face challenges in accurately identifying individuals due to variations in clothing, viewing angles, and other confounding factors, leading to degraded recognition performance, especially when operating on frontal-view walking data which has minimal gait cues.
Innovation Solution
A novel gait identification system utilizing a convolutional neural network (CNN) with autoencoder framework and long short-term memory (LSTM) for disentangling pose and appearance features, generating robust gait features from frontal-view RGB videos, and leveraging a Frontal-View Gait (FVG) dataset to improve recognition accuracy and computational efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional gait recognition methods are used, then the system can identify individuals, but recognition performance degrades due to variations in clothing, viewing angles, and other confounding factors
Solution Approach 1:
The patent segments the gait recognition task into two distinct feature spaces: pose features (capturing body configuration and movement) and appearance features (capturing clothing and visual characteristics). By disentangling these features through separate neural network streams, the system can focus on pose features for identification while being invariant to appearance variations, thereby resolving the contradiction between recognition accuracy and sensitivity to appearance changes.
Solution Approach 2:
The patent introduces an intermediary reconstruction module that learns to reconstruct appearance features from pose features. This intermediary mechanism enables the system to understand the relationship between pose and appearance without allowing appearance variations to directly impact the recognition decision, thus maintaining reliability while being robust to harmful appearance factors.
2Productivity
If the system operates on frontal-view walking data, then it can process available video data, but recognition performance degrades due to minimal gait cues in frontal views
Solution Approach 1:
The patent changes the feature extraction parameters by using pose estimation coordinates and body configuration descriptors instead of traditional silhouette-based features. This parameter transformation allows the system to extract meaningful gait information from frontal views by focusing on subtle pose variations that are less apparent in traditional representation spaces, thereby improving measurement precision while maintaining productivity.
3Reliability
If the system uses disentangled representation learning, then it achieves invariance to appearance variations, but computational complexity increases due to multiple neural network streams
Solution Approach 1:
The patent implements multi-functional neural network components that serve multiple purposes: the pose estimation network extracts both pose features for identification and provides input for the appearance reconstruction module; the reconstruction module both learns appearance-pose relationships and enforces disentanglement. This universality reduces overall system complexity while achieving invariance to appearance variations.
Data Source
AI summary
Gait, the walking pattern of individuals, is one of the most important biometrics modalities. Most of the existing gait recognition methods take silhouettes or articulated body models as the gait features. These methods suffer from degraded recognition performance when handling confounding variables, such as clothing, carrying and view angle. To remedy this issue, a novel AutoEncoder framework is presented to explicitly disentangle pose and appearance features from RGB imagery and a long short-term memory integration of pose features over time produces the gait feature.