Gait recognition using description of shape synthesized by planar homography

In the gait recognition, dependency to the walking direction is serious problem because most features obtained from gait sequences for recognition vary with dependent to the walking direction. To extract steady features from the gait sequence in this case, it is noticeable approach to synthesize gait sequences to the canonical-viewed ones. However, even though the synthesized gait is used for the feature extraction, it is required to describe the gait sufficiently for robust recognition. Therefore, the target of this paper is to find a method to reduce the directional dependency, and then apply adequate description for the gait sequences to recognize the gait, which includes a few distortion caused by synthesizing method. To overcome the problem of directional dependency, we propose a synthesis method to compose gait sequences to the canonical-viewed ones based on the planar homography, which is estimated by only using the given gait sequence with simple operation.

The estimated homography by our method is not perfect transformation to make the canonical-viewed gait sequence. Thus, to describe an individual gait sufficiently, we adopt the Shape Sequence Descriptor (SSD), which describes shape information and variation caused by motion, simultaneously. In general, the SSD is used for recognizing motion, which is presented by the fixed object, or person. Thus, it does not be directly applied to the gait recognition because gait sequences is accompanied with positional change, and all of features in the SSD is not significant to the gait recognition. Thus, we modifies the SSD to apply our recognition method, and also, select features according to the significance for recognizing gaits.

From the experiment with real gait sequences, in the restricted condition where the directional dependency has controlled by using the perpendicular gait sequences, the proposed synthesizing method outperforms the method based on simple normalization in the size by about 10 %. In the case using different directional gait sequence, performance of the method using the normalization dropped drastically by 44 % referring to the perpendicular case. It is caused by the effect of directional dependency in the gait sequences. However, the proposed synthesis method improves the performance by about 20 % comparing to the normalization method. From these results, we verify the proposed method can successfully compensate the variation due to the direction of walking and show the reasonable performance of gait recognition.