Abstract:
Gait is an important biometric for identification of people and it is also more difficult to conceal. It is applied in various forensics techniques because the gait of each person varies significantly from others and it is unique to each person. Different studies make use of different covariates for gait identification namely speed,view,cloth, time. These covariates affect the recognition process adversely and decrease the recognition rate. This study makes use of View covariate which poses a significant challenge in the identification process.The angle of the camera changes the features to be learned drastically and pose a challenge to identify a person in different viewing condition The study was done on a data set of very high intra-class variance and it presents an effective way to reduce the variance. The main challenge in the study was the extremely high variation present even among the samples of a single class of the dataset which was dealt with by using a score based algorithm which uses Structural Similarity, Euclidean Distance and Mean squared error. The prominent GEInet was also used which produced a minimal accuracy of 26%.This methodology was applied by training the GEInet from scratch. Various techniques are using pre-trained deep networks for classification and many of these have been highly successful. The study made use of previously trained deep networks such as AlexNet and GoogleNet to classify the highly variate data-set. This method was highly successfuly in producing a markedly high level of positive classification on the gait energy images. The study was successful is producing a mean accuracy of 99% percent in the case of AlexNet and 90% in the case of GoogleNet.
