Robust Gait Recognition Using Adaptive Random Depth Subspace From Depth Information

Researchers proposed a number of approaches for identifying individual by gait. However, there are a number of covariates that affect the performance of gait recognition of individual when appearance-based approaches are applied such as view point and clothing type. The traditional gait recognition using binary images performed best when a side view point image is applied. However, depth information provided by depth sensors, such as Kinect, contains the information of physical distance from the sensor to a human body part at each pixel while walking, and it shows good performance in frontal gait recognition. Meanwhile, related works which focus on recognition of individual using depth data proposed whole-based methods and perform worse when a subject changes his/her clothing type or bag-gage carrying conditions. Since these methods treat each body part equally ignoring the influence of various clothing types and walking postures, and classify the subjects by global descriptor which involves interferences due to various walking conditions. In this study, the author concentrates on exploring the applicability of depth information for robust gait recognition of individuals from frontal viewpoint. The contributions of this thesis is that proposing a part-based approach named adaptive random depth subspace(ARDS) to represent the walking pattern by local descriptors(subspaces). The proposed method is more advantageous than the related works in the following point: 1) The local descriptor is able to avoid the negative effect of covariates(changing clothes and carrying objects) and improves the accuracy in gait recognition of individual. 2) Only normal walking condition(wearing T-shirt, regular pair of pants and natural swinging arms) is assumed in gallery set for references, which is realistic in real-world applications. 3) The proposed method is capable of selecting gait features from each body part adaptively without assuming the walking condition in each test gait sequence. Each subject's gait sequence is defined as a series of static images which are extracted frame-by-frame from the depth gait video in one full gait cycle. In the stage of depth gait representation, the author proposed a gait representation technique, which is motived by `Depth Gradient Oriented Histogram Energy Image(DGHEI)', as a gait template for selecting adaptive random depth subspace. Then, the calculated cell-based matrix is considered as the depth gait representation instead of the depth gait images. In the stage of adaptive random depth subspace framework, the human body region is segmented into 4 body parts. Then, without assuming specific clothes type and baggage carrying condition for each input test sequence, the author selects cells from each body part randomly and combine the selected cells to new gait features, which is called `subspaces'. Finally, in the individual classification stage, a subspace extracted from a test sequence is matched with all subspaces extracted from the sequences in a gallery set. The final decision of classification is taken by majority voting. Experimental results showed the effectiveness compared with other methods using depth information.