Performance analysis of gait recognition with large perspective distortion

In real security scenarios, gait data may be highly distorted due to perspective effects and there may be significant change in appearance, orientation and occlusion between different measurements. To deal with this problem, a new identification technique is proposed by reconstructing 3D models of the walking subject, which are then used to identify subject images from an arbitrary camera. 3D models in one gait cycle are aligned to match silhouettes in a 2D gait cycle by estimating the positions of a 3D and 2D gait cycles in a 3D space. This allows the gait data in a gallery and probe share the same appearance, perspective and occlusion. Generic Fourier Descriptors are used as gait features. The performance is evaluated using a new collected dataset of 17 subjects walking in a narrow walkway. A Correct Classification Rate of 98.8% is achieved. This high recognition rate has still been achieved using a modest number of features. The analysis indicate that the technique can handle truncated gait cycles of different length and is insensitive to noisy silhouettes. However, calibration errors have a negative impact upon recognition performance.