Year: 2022
Type: conference-paper
Abstract: Recognition of human gait in the wild remains a weakly studied problem due to many factors. One of these factors is how to estimate the distance and location of a person in relation to the camera in order to find the best position for visual capturing. Urban video surveillance usually does not use additional sensors, such as RGB-D cameras, LiDAR sensors or other distance detection devices. Therefore, scene depth estimation followed by distance estimation of moving objects is useful for their capturing and tracking before being recognized. A special feature of the gait recognition problem is that the depth map needs to be adjusted only for close distance. This leads to the use of ordinal regression in the model for monocular depth estimation. We offer a depth ordinal regression CNN, output of which is concatenated with the output of a visual deep-based detector. This works as an initial stage before walking style recognition or gait recognition, depending on the distance a person is moving. The experiments were conducted using publicity available gait recognition datasets and own dataset. It has been shown that the use of such an initial procedure increases the accuracy, since moving objects from a large distance are preventively rejected.
Source: Procedia Computer Science
Language: English
Subfield: Biomedical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4313124705