Gait recognition using Local Ternary Pattern (LTP)

Abstract

Local Ternary Pattern (LTP) is usually applied for texture classification problems. In this work, we propose LTP for human gait characterization for the purpose of human identification. Our proposed method is based on the Gait Energy Image (GEI) whereby edge information over a complete gait cycle is extracted. However, GEI does not contain enough human body structure information for human recognition purpose. Therefore, LTP is used to extract texture information from all pixels in the human gait region which preserves more discriminative features of the subject. Gait cycle estimation is computed by using the aspect ratio of the subject's bounding box. After that, LTP features are averaged over a full gait cycle and a 2D joint histogram of the LTP is computed. At the end, K nearest-neighbor (k-NN) is used to obtain the final recognition results. The proposed method achieved higher accuracy compared to other methods when tested on the CMU MoBo human gait database. The proposed LTP method is easy to implement and also has the advantage of significantly lower computation time.