Article Preview
TopIntroduction
In the domain of computer vision, Human Gait Recognition (HGR) is an active research area due to its unobtrusiveness for identification and verification. In comparison with other biometric devices such as iris and face detection (Choudhury and Tjahjadi, 2015; Lishani, Boubchir et al., 2017; Dey, Ashour et al., 2018; Wang, Li et al., 2018), the gait offers an opportunity to recognize human even at a distance point by utilizing a video camera. Existing research in this area has explained that gait is likely to become a robust biometric for assistance in many applications such as surveillance in airports, bus stations, clinical analysis, banks surveillance systems and forensic applications (Aqmar, Fujihara et al., 2014; Yasmin, Sharif et al., 2016; Zeng and Wang, 2016).
Recently, several image processing and machine learning-based techniques are introduced for HGR. These techniques can be separated into two classes such as model based (MB) (Bashir, Xiang, & Gong, 2010) and model free (MF) (Zeng, Wang, & Li, 2014). The MB approach takes structural model of a human body in the absence of motion. Afterwards, parameters of these structural models are further applied as features like joint angles. The major advantage of this approach is to express high-level model of a human body but it is complex because of high cost and computational time. This approach also works better over the view invariant and cofactors like carrying, clothing and shadows which affects the recognition rate (Zeng, Wang et al., 2014, Arora, Hanmandlu et al., 2015). The model free class operates on silhouette images of a human body instead of a structural model. This approach is less sensitive for the quality of an image as compared to model-based class. Also, it has low computational time and cost. The major challenge of this approach is sensitivity because of several problems such as carrying, clothing and shadows (Piccardi 2004, Rida, Jiang et al., 2016; Wu, Huang et al., 2017).
A general gait recognition framework consists of several sub steps such as preprocessing, ROI detection, feature extraction and recognition. The preprocessing step has much importance for obtaining high recognition rate because the raw input videos have a complex background and low contrast. The low contrast videos affect ROI detection which later on degrades the recognition accuracy. Therefore, preprocessing is a major step for improving the contrast of input video and also removes extra noise such as change in background, variations and human change. In gait recognition, different segmentation methods have been introduced for ROI detection such as background subtraction, thresholding, watershed and few more (Piccardi, 2004; Gupta, Dixit et al., 2014). There are several types of features which exist in computer vision such as texture, color, geometric, shape, Gabor and wavelet transform. These features are high in number of dimensions; therefore, they reduce the accuracy. This kind of problem is resolved by several researchers by the implementation of reduction techniques. The major reduction techniques are Principal Component Analysis (PCA) (Ryu and Kamata, 2011), ICA, genetic algorithm and few more (Khan, Sharif et al., 2016). Finally, the reduced features are utilized by supervised learning methods as SVM, Fine K-Nearest Neighbor (FKNN), neural network, decision trees and regression models (Abdullah and El-Alfy, 2015; Khan, Sharif et al., 2016; Nida, Sharif et al., 2016). These methods perform significantly well when the extracted number of features is unique and have no redundancy between them. Therefore, classification accuracy fully depends on the extracted number of features.