Chin-Pan Huang

Human Action Recognition Using Histogram of Oriented Gradient of Motion History Image

This paper presents a human action recognition method using histogram of oriented gradient (HOG) of motion history image (MHI). First, the proposed method generates MHI with differential images which are obtained by frame difference of successive frames of a video. The histogram of oriented gradient (HOG) of the MHI is then computed. Finally, support vector machine (SVM) is applied to train an action classifier using the HOG features. We discovered that the new method improves recognition rate significantly. Moreover, our algorithm does not require the generation process of human silhouette, and therefore the performance is also increased. Experimental results are provided to show the promising performance of the proposed method.

Progressive sharing for a secret image

Based on the wavelet transform, a new progressive sharing scheme is proposed to share a secret image into several shadow images using SPIHT encoding processes and Shamirs threshold scheme. Quality refinement of the recovered image is achieved by the data consumed from the threshold number (r) of shadow images and each single shadow image reveals no information about the secret image. The size of each shadow image is smaller than 1/r of the secret image and any number of shadow images that is less than r reveals no information about the secret image. The proposed approach is secure for image sharing and provides excellent peak signal-to-noise ratio (PSNR) versus rate performance. Experimental results have demonstrated the promising performance of this method in progressive sharing.

An adaptive eye gaze tracker system in the integrated cloud computing and mobile device

This paper proposed an adaptive method for tracking eye gaze in the integrated cloud computing and mobile device environment. The task begins with extracting the eye position and the iris contour base on geometrical features. These local gaze features are calculate and integrated to train a neural network. And the estimated gaze point is outputted from the trained NN (Neural Network) in the cloud computing. A utility function is proposed to decide the functionality is performed in the cloud or mobile device adaptively based on device and network conditions. Besides, our proposed method can improve system performance as well as overcome the problem for limited resource of mobile device.

Detection and Recognition of Scoreboard for Baseball Videos

This paper presents an effective and efficient detection and recognition of scoreboard caption method for baseball videos. The method first identifies the scoreboard type using template matching and then extracts the caption region of each type. Next it recognizes the extracted caption utilizing a novel digit recognition scheme which is constructed by a simple neural network classifier. It results in a much simpler method with significantly higher recognition rate over that of the universal OCR scheme. Experimental results demonstrate the effectiveness of the proposed method and indicate that it identifies twelve scoreboard types correctly and recognizes scoreboard caption over 98%.

Human Action Recognition Using Depth Images

This paper presents a human action recognition algorithm using a depth image. First, 3D coordinates of the body’s joints of each frame are generated from the depth image. Then, the proposed method applies normalization and quantization processes to the body joints of all frames of the action video to obtain a 3D histogram. The histogram is projected onto xy, xz, and yz plans sequentially and combined into a one-dimensional feature vector. For dimension reduction, the principal component analysis (PCA) technique is applied to the feature vector to generate an action descriptor. To further improve the recognition performance, a decision tree method is developed to divide input actions into four main categories. The action description vectors of each category are used to design its respective support vector machine (SVM) classifier. Each SVM classifies the actions of a category into one type of actions. Experimental results verify that our approach effectively rules out the interference of background and improves the recognition rate.

A Vision-based System to Help Senior Citizens for Memory Recall of Object-placing

Human memory ability is gradually degenerated followed by the increasing of age. It is annoying for senior citizens to often forgetting and trying to recall the places of those objects commonly used in their daily lives. This paper presents a preliminary study for developing a vision-based system and helping senior citizens to recall the process of object-placing. In the initial stage, SURF (Speeded Up Robust Features) is adopted to extract feature points from the object image that are further used for the matching to its corresponding video clips in the video database. Experimental results have demonstrated that correct video clips can be retrieved from the database and played for the user to recall after the desired object image is provided. Although impressive results have been achieved, to increase the robustness of this system, more features and objects will be analyzed and used in the future.

Novel high-capacity data hiding technique using summation and LSD parity of pixel groups

The aims of designing algorithms for data hiding applications are to increase the hiding capacity and maintain the original image quality. However, both goals are often contradicted to each other. Owing to this, the main idea of this paper is to embed the secret bit sequence into pixel groups by using the summation of neighboring pixel values and the achieved parity of Least Significant Digit (LSD). The summation method utilizes the concept of pixel groups from the method of pixel value differencing (PVD) but not using the difference values for data hiding. Instead, the decimal representation of LSD summation relationship from two neighboring pixels is adopted for adjusting pixel values for data hiding. At the same time, the parity of the second pixel value is used for increasing one more hidden-bit while using the summation relationship for data hiding. Data extraction can be done in the reverse direction. Therefore, the overall hiding capacity can be greatly improved. Experimental results show that the proposed data hiding algorithm can have better performance than normal PVD methods. Also, the better image quality can be assured.

A Vision-Based Universal Control System for Home Appliances

Based on computer vision techniques, this paper presents a universal remote control system for home appliances. This system consists of three major components: a paper control panel, a web camera wore on the user’s chest, and a laptop computer. The system operates as follows. First, the user points his finger tip to a virtual button on the paper panel to select a specific appliance. Second, a function button is pointed at to operate an assigned function. The user’s hand image is captured by the camera and the virtual button pointed by the fingertip is detected by computer vision techniques. Then a specific infrared code is emitted to control the corresponding appliance for a specific function. The advantage of this universal remote control system is that several remote controllers can be integrated together to simplify the operations of different appliances. The system performance is shown in the experimental results.

A multiple description coding method based on 3-D SPIHT for videos

This study presents a new multiple description coding method based on a three-dimensional set partitioning in hierarchical trees (3-D SPIHT) algorithm in WiMAX for videos. Applying a poly-phase sampling technique to sample an input video over its intra-frame or inter-frame, several independent and highly correlated descriptions are generated. Then, a 3-D SPIHT algorithm is applied to each of the produced descriptions. Finally, the coding results are transmitted via different channels. This proposed approach achieves high compression ratio, enhances transmission speed, and more importantly, attains acceptable video quality when channel congestion occurs. Experimental results are provided to demonstrate promising performance of the proposed method.

Object Recognition Based on Fuzzy Morphological Polynomial Signal Representation

Fuzzy morphological polynomial (FMP) signal representation provides the merits of compact representation and low computation complexity. Based on FMP representation, a theoretical study for object recognition is presented in this paper. The study indicates that superior properties of FMP can be used to develop an object recognition system. Examples are given to illustrate the effectiveness of FMP representation for object recognition.