Hiromitsu Hama

Fusion of Infrared and Visible Images for Robust Person Detection

In the current context of increased surveillance and security, more sophisticated and robust surveillance systems are needed. One idea relies on the use of pairs of video (visible spectrum) and thermal infrared (IR) cameras located around premises of interest. To automate the system, a robust person detection algorithm and the development of an efficient technique enabling the fusion of the information provided by the two sensors becomes necessary and these are described in this chapter. Recently, multi-sensor based image fusion system is a challenging task and fundamental to several modern day image processing applications, such as security systems, defence applications, and intelligent machines. Image fusion techniques have been actively investigated and have wide application in various fields. It is often a vital pre-processing procedure to many computer vision and image processing tasks which are dependent on the acquisition of imaging data via sensors, such as IR and visible. One such task is that of human detection. To detect humans with an artificial system is difficult for a number of reasons as shown in Figure 1 (Gavrila, 2001). The main challenge for a vision-based pedestrian detector is the high degree of variability with the human appearance due to articulated motion, body size, partial occlusion, inconsistent cloth texture, highly cluttered backgrounds and changing lighting conditions.

A Markov Random Walk Model for Loitering People Detection

Today video surveillance systems are widely used in public spaces, such as train stations or airports, to enhance security. In order to observe large and complex facilities a huge amount of cameras is required. These create a massive amount of data to be analyzed. It is therefore crucial to support human security staff with automatic surveillance applications, which will create an alert if security relevant events are detected. This way video surveillance could be used to prevent potentially dangerous situations, instead of just being used as forensic instrument, to analyze an event after it happened. In this treatise we present a surveillance system which supports human operators, by automatically detecting loitering people. Usually, loitering human behavior often leads to abnormal situations, like suspected drug-dealing activity, bank robbery, and pickpocket, etc. Thus, the problem of loitering detection in image sequences involving situations with multiple objects is studied based two dimensional Markov random walks in which both motion and appearance features describing the movements of a varying number of objects as well as their entries and exits are used. To obtain efficient and compact representations we encode the spatiotemporal information of intra-inter trajectory contexts into the transition matrix of a Markov Random Walk, and then extract its stationary distribution and boundary crossing probabilities as final detection criteria. The model is also made less sensitive to uninteresting objects occluding the region of interest by integration out their effect on the observation probabilities. The resulting system is tested on the real life dataset scenarios giving 95% performance results.

Dominant Color Embedded Markov Chain Model for Object Image Retrieval

By focusing on the decoding process to improve coding efficiency, we propose an effective video coding method by employing parameter estimation in the decoding process. A bitrate reduction can be achieved when the proposed method is applied to the DC transform coefficient and the motion vector of H.264. 0.25%-0.84%.

Optimal Crawling Strategies for Multimedia Search Engines

In this paper we propose a novel optimal crawling strategy for next-generation multimedia search engines. We consider here a Web crawl as a two-dimensional (2D) random walker on a graph whose vertices are the Web pages and whose edges are the hyperlinks. The proposed crawler is a two-part scheme optimizing the crawling process in such a way that the average level of staleness over all pages is minimized and the quality of search engine from user’s perspective is maximized. In doing so, we employ techniques from probability theory and the theory of functional equations which are highly computationally efficient-crucial for practicality because the size of the problem in the Web environment is immense. We show that a combination of breadth-depth crawling including the largest sites is a practical and optimal strategy. In particular, several probabilistic models for user browsing in infinite Web are proposed and studied to estimate how deep and breadth a crawler must go to download a significant portion of the Web site that is actually visited. Experimental and simulation results show that a crawler needs to download just a few levels in depth and breadth to reach the maximum number of pages that users actually visit. It also suggests that the largest sites should be included in the crawling process.

Detection of Tire-Road Contact Point for Vehicle Position Estimate Considering Shape Distortion in a Circular Fisheye Image

In this paper, we propose a new method for detecting tire-road contact points and estimating the position of a vehicle on a road taking the distortion in a circular fisheye image into consideration. In our proposed method, we use the distortional parameter of two concentric circles composed of the inner wheel and the outer tire by considering projection system of fisheye lens, and the searching the gray scale profile in the distortional direction are derived from each pixel that is outer the outline of the wheel region. And it is confirmed that the proposed method is effective in circular fisheye image by experimental results.

Background modeling using special type of Markov Chain

Background modeling is important in video surveillance for extracting foreground regions from a complex environment. In this paper, we present a novel background modeling technique based on a special type of Markov Chain. The method is a substantial extension to the existing background subtraction techniques. First, a background pixel is statistically modeled by a linear regressive Gamma Markov distribution. Then, these statistical estimates are used as important parameters in background update schemes. The experimental results show that the proposed model is less sensitive to movements of the texture background and more robust for real time segmenting the foreground object accurately.

Bundling Multislit-HOG Features of Near Infrared Images for Pedestrian Detection

In this paper we present a novel scheme where image features are bundled into local groups. Specifically, features of Near Infrared (NIR) images extracted by using Histogram of Oriented Gradients (HOG) descriptor and those by our multislit method are bundled into a single descriptor. The method involves first localizing the spatial layout of body parts (head, torso, and legs) in individual frames using multislit structures, and associating these through a series of extracting HOG features. A bundled feature vector describing various types of poses is then constructed and used for detecting the pedestrians. Experiments with a database of NIR images show that our scheme achieves a substantial improvement in average precision over the baseline conventional HOG approach. Detection and recognition performance is less computationally expensive than existing approaches.

Accurate Estimation of Wheel Center Points for Estimate of Vehicle Baseline Length in a Circular Fisheye Image

Vehicle classification is very important in traffic flow analysis for traffic monitoring systems. Accurate estimation of wheel baseline length can provide a useful cue for vehicle classification. In this paper, we propose a new method for detecting accurate wheel center points and estimating the wheel baseline length of a vehicle taking the distortion in the circular fisheye images into consideration. Our proposed method consists of two steps, considering distortional shape of the wheel in a circular fisheye image. At the first step, tire-road contact points are roughly detected based on the centroid of wheel region. At the second step, the wheel area is calculated based on detected contact points, and the wheels center points are accurately detected by position matching of wheel area. And it is confirmed that the proposed method is effective in circular fisheye image by experimental results.

Reliability Based Web Information Ranking System

In this paper, we propose a reliability based Web information ranking system which enables searching useful and reliable knowledge information. The proposed system will contain subsystems for reliability ranking, information clustering based on reliability. The reliability ranking system will estimate the likelihood that a statement on the Web can be trusted using standards developed by information scientists, and the link structure of associated Web pages. The clustering will cluster relevant and reliable information based on whether or not they can be trusted or not. We test these models on an academic search engine and show how the reliability information ranks can be used as a useful knowledge.

An Optimal Choice of Morphological Operating Center for Object Image Retrieval

In this paper we introduce a novel and simple schemes to develop an optimal choice of morphological Operating Center (OC) for object image retrieval. A variation of the standard morphological operators which require the choice of an OC is discussed. The proposed method is based on combinations of statistical and dynamic programming techniques in which recursive equations on the basis of dilation by using the principle of optimality and minimizing unnecessary background area of the objects are applied. We also present the application of mathematical morphology with Structuring Elements (SEs) which are elongated in the angular direction. The experimental results show that the optimal choice of OC provides satisfying retrieval results.