Research Of Threshold Segmentation Algorithms And Pedestrians Detection In Infrared Image

Image thresholding is an important method of image processing, which is applied in many fields such as motion estimation and object detection recognition. Human is the most important object in image segmentation based applications, and human detection is a hot and difficulty issue in the field of computer vision, which is applied in many fields such as video surveillance, driving assistance. Due to human objects are non-rigid objects, as well as the high degree of intra and inter variability of human in the visual domain, it is difficult to detect human. In visible light, due to the changes of background in monitor scene with the light changes of surroundings, as well as much interference of motion objects in the scene, it is difficult to distinguish object pixels from background pixels. With the fast development of infrared technology, for complex background can not influence the target detecting in infrared image, infrared devices have been an important technology of pedestrian detection. So the research of image thresholding and pedestrian detection possesses important theoretical meaning and extensive application value. The paper includes the following important sections:1. The basic principle of image thresholding and some usual methods are researched and analyzed. An optimal evolution algorithm for image thresholding is proposed. According to the algorithm, image thresholding become an optimization problem. To deal with situation of bimodal histogram and multimodal histogram, the algorithm distinguishes object from background well. Based on the thesis that the optimal threshold is the best direction of biological evolution, the updating model of evolution direction is established.2. In the stage of infrared image preprocessing, for infrared images have low signal-to-noise ratio and contrast ratio, the median filter is applied, in which image noise pixels are marked based on the number. While the image noise pixels are filtered, the detail of image can be reserved, also the image resolution can be impoved.3. In the stage of infrared image thresholding processing, an improved image thresholding method based on single Gaussian model is proposed and compared with other methods of image thresholding by experiments. Due to the pixel gray value in infrared images can be approximately modeled by a single Gaussian distribution, this method can detect highlight pixels and obtain regions of interest in images.4. In the stage of infrared image post-processing, as to morphological methods the binary image is de-noised, as well as the connected domain is marked. Due to the principle of area merging and deleting selectively, which is proposed based on the characteristics of human region, excessive merging is controlled, and the number of false positives is reduced. Also the redundant regions of interest are reduced.5. In the stage of pedestrian recognization in infrared image, the initial classifier based on 2D histogram template and the fine classifier based on HOG describer are designed. These classifiers are trained by the method of FLD and linear SVM respectively. Then the samples of ROI are classified and marked. The experimental results show that this algorithm has high true positive rate and low false positive rate, can determine whether one ROI is a person or not effectively.