Gait recognition for human identification and re-identification using deep neural network

Authorities are always looking for tools and require a more effective method of tracking and re-identifying a person of interest. Gait is the unique pattern during humans' walking and receives more attention due to its ability to continually collect samples from afar without the subject's awareness seamlessly. When integrated with machine learning, gait biometrics provide powerful tools for security or authentication processes.Early gait recognition works are based on fixed direction walks (view-dependent), which need to be operated in view-dependent scenarios, and are not suitable for freestyle walks. This research proposed the gait recognition techniques for freestyle walks (view-independent) on structural gait data (model-based). We present three novel gait recognition techniques:(1) We propose an improved data preprocessing technique. The proposed technique focuses on frames with high joint tracking status and improves their quality by using averaging to smooth the walking sequence and reduce noise. We used gait features that do not depend on the viewpoint, such as limb lengths, the angles from a group of joints, etc. The technique was called "double-window". (2) We propose new ways to create a gait feature called joint replacement coordinates (JRCs) and use them with a new CNN design. JRCs reduce noise in body posture and describe body movements using movement information from three connected joints. The body posture's movement is developed from every group of joints over the body. The proposed technique is called JRC-CNN gait recognition. (3) We propose a new unsupervised gait recognition technique based on a deep metric learning model. Most of the existing gait techniques are identified an unknown identity from a database of known identity. The proposed technique is called "regional-LSTM". This is a new concept to extract the rhythm of movements in different body regions by creating a separate LSTM model for each region. The regional-LSTM learning model is a representation function that transforms gait characteristics to an embedding space with a low degree of intra-class similarity and a high degree of inter-class similarity, where its similarity can be easily measured by using the ordinary distance functions.