An RGB Gait Anonymization Model for Low-Quality Silhouettes

Gait anonymization while maintaining naturalness is used for protecting a person's identity against gait recognition systems when a video of the person walking is uploaded to social media. There has been some research on gait anonymization, but only for high-quality silhouette gaits. We present an RGB gait anonymization model for low-quality silhouette gaits that can generate natural, seamless anonymized gaits for which the original silhouettes cannot be extracted correctly. Our model includes two main networks. The first one, a deep convolutional generative adversarial network, is used to anonymize the original gait by adding to it a random noise vector. By training on high-quality silhouette data, this network can generate a high-quality anonymized silhouette sequence from a low-quality silhouette one. Restricting its input to binary silhouette sequences instead of color gaits forces it to focus on anonymizing the gait rather than changing body color. The second main network, which follows the first one, colorizes the anonymized silhouette sequence generated by the first network by using the color of the original gait. Evaluation in terms of success rate and naturalness demonstrated that our model can anonymize gaits while maintaining naturalness.