Deep Trajectory Based Gait Recognition for Human Re-identification

The popular techniques of gait recognition rely on the appearance information, such as Gait Energy Image (GEI). However, they need the pre-processing stage of silhouette segmentation in a walking video. This may not be efficient when the complete silhouette could not be obtained under the cluttered walking environment. It is also sensitive to the changes of walking conditions. Thus, this paper comes up with a new solution using the dense trajectory. This technique is commonly used in the action recognition domain. In this paper, it is used to extract the gait information. The key points and their corresponding trajectories are detected. Then, HOG, HOF, MBHx, MBHy and dense trajectory are extracted from each key point as the point descriptor. In the training phase, the bag of word (BoW) are trained using the extracted point descriptors from the training gait videos. Finally, in the testing phase, the BoW is extracted for each gait video, as the gait feature. The experimental result based on the well-known CASIA gait database B shows the promising performance of the proposed method, under various views.