A practical technique for gait recognition on curved and straight trajectories

Many studies show the effectiveness of gait in surveillance and access control scenarios. However, appearance changes due to walking direction changes impose a challenge for gait recognition techniques that assume people only walk in a straight line. In this paper, the effect of walking along straight and curved path is studied by proposing a practical technique which is based on the three key frames in the start, middle and end of the gait cycle. The position of these frames is estimated in 3D space which is then used to estimate the local walking direction in the first and second part of the cycle. The technique used 3D volume sequences of the people to adapt to changes in the walking direction. The performance is evaluated using a newly collected dataset and the Kyushu University 4D Gait Dataset, containing people walking in straight lines and curves. With the proposed technique, we obtain a correct classification rate of 98% for matching straight with straight walking and 81% for matching straight with curved walking averaged over both datasets. The variation in walking patterns when a person walks along a straight or curved path is most likely to be responsible for the difference. In support of this, the recognition rate when matching curved with curved walking is 99% on our dataset.