PGcGAN: Pathological Gait-Conditioned GAN for Human Gait Synthesis

Pathological gait analysis is constrained by limited and variable clinical datasets, which restrict the modeling of diverse gait impairments. To address this challenge, we propose a Pathological Gait-conditioned Generative Adversarial Network (PGcGAN) that synthesises pathology-specific gait sequences directly from observed 3D pose keypoint trajectories data. The framework incorporates one-hot encoded pathology labels within both the generator and discriminator, enabling controlled synthesis across six gait categories. The generator adopts a conditional autoencoder architecture trained with adversarial and reconstruction objectives to preserve structural and temporal gait characteristics. Experiments on the Pathological Gait Dataset demonstrate strong alignment between real and synthetic sequences through PCA and t-SNE analyses, visual kinematic inspection, and downstream classification tasks. Augmenting real data with synthetic sequences improved pathological gait recognition across GRU, LSTM, and CNN models, indicating that pathology-conditioned gait synthesis can effectively support data augmentation in pathological gait analysis.

💡 Research Summary

The paper introduces PGcGAN, a Pathological Gait‑conditioned Generative Adversarial Network designed to synthesize realistic 3‑D joint‑keypoint gait sequences for six distinct pathological gait categories. The authors motivate the work by highlighting the scarcity and high variability of clinical gait datasets, which hampers the development of robust machine‑learning models for diagnosis and rehabilitation. PGcGAN addresses this gap by embedding one‑hot encoded pathology labels into both the generator and discriminator, thereby enabling explicit control over the type of gait being generated.

The generator follows a conditional auto‑encoder architecture. A Gaussian noise tensor is first encoded into a latent vector via a temporal convolutional encoder. This latent representation is concatenated with the pathology label and fed into a temporal convolutional decoder that reconstructs a full‑length gait sequence. Reconstruction is enforced through an L2 loss between the generated sequence and the real input, preserving biomechanical structure such as joint‑to‑joint distances and temporal coherence. The discriminator receives either a real or synthetic sequence, each paired with the same pathology label, and classifies them using stacked 1‑D temporal convolutions followed by fully‑connected layers with spectral normalization for training stability.

Training optimizes a weighted sum of the standard adversarial loss and the reconstruction loss, with hyper‑parameters λ_adv and λ_rec balancing realism against structural fidelity. The model is trained on the Pathological Gait Dataset (Kinect‑V2 captured 3‑D poses), filtered to retain sequences longer than 60 frames, resulting in 7,157 samples.

Evaluation proceeds along three axes. First, dimensionality‑reduction visualisations (PCA and t‑SNE) show that synthetic samples occupy the same latent clusters as real data, indicating that the generator captures the underlying distribution of each pathology. Second, a biomechanical validation compares joint‑angle trajectories and reports an R² of 0.94 between real and synthetic normal‑gait trajectories, confirming high kinematic fidelity. Third, the authors assess the utility of synthetic data for downstream classification. Three deep‑learning classifiers (GRU, LSTM, CNN) are trained under three conditions: (a) real data only, (b) synthetic data only, and (c) a mixture of real and synthetic data. While synthetic‑only training yields lower accuracies (e.g., GRU 87.65 % vs. 91.87 % with real data), augmenting the real set with synthetic samples consistently improves performance (GRU rises to 92.61 %, CNN to 89.56 %). This demonstrates that PGcGAN‑generated gait sequences can enrich the training distribution and help models generalise to unseen pathological variations.

The paper also contrasts PGcGAN with prior synthetic gait works that focus on normal gait or specific disorders (e.g., FoGGAN for freezing‑of‑gait, Morph for physics‑based motion). Unlike those, PGcGAN simultaneously handles multiple pathologies and validates both statistical alignment and biomechanical plausibility, achieving superior quantitative metrics.

Limitations are acknowledged: the one‑hot label scheme does not capture continuous severity scales, the latent space remains opaque, and synthetic‑only models underperform, suggesting that generated data do not yet fully replace real recordings. Future directions include exploring continuous conditioning, disentangled latent representations, and hybrid physics‑GAN frameworks to further improve realism and controllability.

In summary, PGcGAN represents a novel contribution to the field of pathological gait analysis by providing a controllable, high‑fidelity data‑augmentation tool that can alleviate dataset scarcity, improve classification performance, and open new avenues for simulation‑based rehabilitation research.