Deep Learning Based Human Locomotion Recognition in Video Sequences

The paper proposes a classification method of human locomotion types from video sequences based on machine learning methods. The purpose is to discriminate between three locomotion types: walking, jogging and running, using only postural features. It is obvious that the result of the classification may be wrong for some frames due to the similarity of some transition postures. For this reason, a temporal filtering procedure is applied for smoothing the results and makes the final decision for the entire video sequence. For some video sequences, the obtained result is different from the a priori known movement type. However, a careful analysis shows that the dynamics of the movement is closer to that determined by classification than the known one. Therefore, the procedure can be applied to correct the movement in sports activities or in medical recovery.