Human Recognition At A Distance In Video PDF Download

In recent years, off-the-shelf cameras became vastly available, producing a huge amount of content that can be used in various applications. Among the applications, visual surveillance receives a great deal of interest. In visual surveillance, for the purpose of advanced security and monitoring system, human recognition at a distance is of importance. Interest in finding specific individual becomes important such as the case in locating missing person and identification of terrorist from recording of video surveillance in public places. Many methods published so far for human recognition in video image analysis focused on face and gait. However, these methods have low recognition performance due to some external factors such as face are prone to fake facial images and gait is susceptible to occlusion. The presence of the occlusion introduces errors into many existing vision algorithms which have yet to be resolved. In such situations vision techniques to identify a human fails because descriptors of part of the human shape may not have any resemblance with the descriptors of the entire human shape. To resolve this problem, an intelligent system for human recognition under partial occlusion is proposed in this study. The hypothesis of the study is that special features of human body shape can be used to identify a person identity from a distance. Each person has different body features characteristics and hence recognizable using these special body shape features. The body shape features used are the head, shoulder, and trunk. The features can be recognized using fuzzy logic (FL) approach and used as inputs to a recognition system based on a multilayer neural network (MNN). For the developed human recognition algorithm, database has been implemented to provide efficiency in displaying, storing, and retrieval of data. Experimental results show that the developed human recognition system is capable of detecting and recognizing human subject from a distance with 98.1% and 77.5% accuracy respectively. For successful identity recognition, the average percentage coverage of occlusion for Single Direction Partial Occlusion Test (SDPOT) from the top, bottom, left, and right of specific human subjects are 0.41%, 10.41%, 8.44%, and 12.27% respectively. Whereas for Multiple Direction Partial Occlusion Test (MDPOT), an average of 40.15% coverage of occlusions is achieved with successful recognition while the lowest is found to be 11.92% and highest is 60.92%.