Real Time Motion Detection for Fast Human Identification Based on Face Recognition

The problem of real time in motion human identification based on face recognition has attracted the attention of the scientists and researchers. Because of real time face recognition in non-cooperative environments is a complex task, more tools and algorithms are needed in order to reduce the processing time required for a complete system to operate in an appropriate way. In this paper is presented an algorithm for motion detection that has low complexity and it’s designed to operate in real time applications. The algorithm presented was designed specially for a human identification based on face recognition. The algorithms consist in low level digital image processing tools, like binarization, 2D filtering, dilate and erode operators. An auto-adaptive background subtraction based on motion energy method is also presented. The algorithm performs very well on real time applications developed with high-level programming language. 1 INTRODUCCIÓN Nowadays there is motion detection and analysis methods, which have many applications, like object tracking through unobserved regions [1], automatic pedestrian counting [2], human tracking [3], non-rigid object tracking [4], motion detection on a desktop interface [5], circular motion analysis [6], reactive motion analysis [7], motion analysis for intelligent transportation [8], and also for image reconstruction [9]. The applications mentioned previously utilize motion detection or analysis techniques. Some of the existing methods for motion detection vary depending on its application, as an example, in [10] is shown a method to detect motion using motion coherence, other utilize code arrays with correlation and convolution properties [11]. Other authors, for their applications, utilize the background extraction technique [12-16]. Neural or Bayesian networks can be utilized as well [1718]. Probabilistic methods are suggested in [19], the frame differencing technique is also well utilized, this technique is also called image difference [20-21]; other algorithms work dividing a video sequence in small subblocks for independent analysis [13] [22-23], other algorithms were designed for mobile cameras [18] [24], in [14-15] Gaussian techniques were utilized but they are computationally expensive. Some of these methods are non-causal systems, they depend on further inputs of a signal, in terms of digital video this means that the video is already recorded or the video is delayed on purpose, for this reason real time implementation of the algorithms is not possible [23] [12]. Real time face recognition systems require being dynamic and dealing with many other processes. To be dynamic means an exhaustive facial search on every single frame when motion is detected on a scene. After an exhaustive research that there is only one specialized motion detection technique for human identification based on face recognition [25], but this algorithm can be improved. Therefore, the main task is to find an appropriate algorithm for motion detection. CIECE – CITEE 2006 Ponencia presentada en el XVI Congreso Interuniversitario de Electrónica, Computación y Eléctrica – II Congreso de Innovación Tecnológica en Eléctrica y Electrónica Cd. Obregón, Sonora; del 5 al 7 de Abril de 2006 172 CIECE CITEE 2006 The algorithm must fulfill the following 3 requirements; first, the algorithm must treat the video as causal signal in order to be implemented on a real time system. Second, the method must avoid computational complexity because there are other main tasks like face detection, feature extraction, and face recognition, not addressed in this paper. Third, one of the outputs of the system must include the image of the object in motion in order to be processed searching a human face in a further work. Some of this metrics were taken from [25-26]. 2 MOTION DETECTION In real time face recognition systems the process of the entire system begins with video acquisition followed by a motion detection and segmentation sub-process. After the segmentation of the object in motion, the sub-process of face detection is activated to extract the facial image if detected. When facial image is available, characteristics are extracted from the facial image, and finally the sub-process of face recognition classifies the identity of the input face. This process is described graphically in Figure 1. Figure 1. Real time face recognition system. Motion detection sub-system stages. In figure is shown the stages of the motion detection sub-system proposed, and will be explained in the following sub-sections. For this project a standard gray scale camera and a standard acquisition card via RCA were utilized for video acquisition. Most of this work is based on [25] which is a previous research focused on evaluate and find an optimal motion detection methods for face recognition applications. 2.1. Image difference. The image difference is a technique that aims to show any change between two images. Frame differencing is also utilized for motion detection, but for face recognition systems has poor results [25]. This method is also known as frame differencing or image differencing and it is widely utilized because of its easier comprehension and implementation. It is described as follows ) 1 , , ( ) , , ( ) , , ( − − = t y x I t y x I t y x Idiff (1) were ) , , ( t y x Idiff denotes the result of subtract the current image ) , , ( t y x I and the previous image ) 1 , , ( − t y x I . The frame reference is t . The Figure 2 show the results of applying the image difference on two images captured in sequence while a person is walking at normal speed. In a) ) , , ( t y x I is shown, and b) shows the result ) , , ( t y x Idiff .