چكيده به لاتين
Identification of people at distance due to its various applications has been of great interest to the researchers recently. In this regards, the rhythm of walking is known as a biometric trait that can be recognized at distance. Most of the algorithms, extract the silhouettes, compute the period of walking and average the silhouettes over a period. The resulting template called Gait Energy Image (GEI) is found as an efficient feature for gait recognition.
Although the GEI feature provide a robust representation tool against noises and failures, it suffers from at least three problems: 1) losing the temporal ordering of gait during the averaging process, 2) applying imprecise human’s motion model, and 3) existing redundant information in final template. To handle these issues, many spatio-temporal approaches has been developed during the last decade. But there is no accurate method to address three above problems in different gait conditions. In this research, a proper spatio-temporal filtering has been proposed to preserve time-ordering of gait in final template. More precisely, the responses of filtering provide the information on orientation and style of human’s motion in each frame of walking. By averaging the responses of such filtering over a gait period, the proposed salient template has been derived accordingly. Furthermore, the weighted local patch method, called weighted Gabor Energy Image (wGbEI), has been proposed to remove redundant and noisy information in final template. The provided features have capabilities to represent the gait in normal conditions, wearing different clothing, shoes, bagging and trivial camera viewing conditions.
In order to evaluate proposed methods, we perform complete evaluations on three well-known datasets. The accuracy of our method, combined with Random Subspace Classifier (RSM), in the USF Human-ID Dataset is 74.46% in Rank1 and 86.71% in Rank5. These rates have been improved by about 4% in Rank1 compared with state-of-the-art methods. Moreover, the proposed biometric system has the error of FAR=38/1000 and FRR=23% at Rank1. In addition, the computational speed of proposed system is 5.5 frame per second which requires 3.6 GBytes memory. The evaluation results indicate that the proposed system is superior in minimizing the errors and has the better accuracies within the dataset with a few computational overheads.
