Gait Recognition using 3D View-Transformation Model
Sprache des Vortragstitels:
Sprache des Tagungstitel:
When it comes to visual based gait recognition, one of the biggest problems is the variance introduced by different camera viewing angles. We generate 3D human models from single RGB person image frames, rotate these 3D models into the side view, and compute gait features used to train a convolutional neural network to recognize people based on their gait information. In our experiment we compare our approach with a method that recognizes people under different viewing angles and show that even for low-resolution input images, the applied view-transformation 1) preserves enough gait information for recognition purposes and 2) produces recognition accuracies just as high without requiring samples from each viewing angle. We believe our approach will produce even better results for higher resolution input images. As far as we know, this is the first appearance-based method that recreates 3D human models using only single RGB images to tackle the viewing-angle problem in gait recognition.