Luo-Wei Tsai

Lane detection using directional random walks

This paper proposes a novel lane detection method for extracting various lane lines from videos using the concept of directional random walks. Two major components are included in this method, i.e., lane segmentation extraction and edge linking. At first, we define proper structure elements to extract different lane mark features from input frames using a novel morphology-based approach. Then, a novel linking technique is proposed to link all ldquodesiredrdquo lane mark features for lane line detection. The technique considers the linking process as a directional random walk which constructs a Markov probability matrix for measuring the direction relationships between lane segments. Then, from the matrix of transition probability, the correct locations of all lane lines can be decided and found from videos. Without defining any mathematical curve models, various road lane shapes and types can be well extracted from road frames even with complicated backgrounds. Experimental results show that the proposed scheme is powerful in lane detection.

Human Action Recognition Using Star Templates and Delaunay Triangulation

This paper presents a human action recognition system for recognizing various behaviors directly from videos. Firstly, we triangulate the human body to different triangle meshes. Then, we use a depth-first search (dfs) scheme to find a spanning tree from the set of meshes. All leafs of the spanning tree are adopted as the extremities. Different from traditional approaches to find the extremities on the targetpsilas silhouette as skeletons, the extremities found from the internal centroids of triangle meshes can represent a human posture more accurately and robustly. To model each human action, all the input skeleton sequences are then transformed into symbol sequences. Then, we design a string matching scheme to measure the similarity between any two human behaviors. Since 2D postures are used in this paper, the above scheme is sensitive to different view points. To solve the view independent problem, a 2D matrix is then constructed for recording the symbol relations between two viewpoints. Thus, our proposed matching scheme is almost view-invariant. Experimental results show that the proposed scheme is a robust, efficient, and promising tool in human action recognition.

Robust face recognition under illumination and facial expression variations

Illumination and expression variations are still a challenging problem in face recognition. In this work, we present an efficient face recognition method which can solve the above two problems with single training sample. At first, the effect of the lighting variation is effectively eliminated by the Mutil-Scale Retinex algorithm. The Active Appearance Model is adopted to extract the facial block feature to establish the component-based face recognition system. Different from other methods which construct the various classifiers corresponding to the specific facial expression, the proposed method decreases the weights of some dominated facial features which are affected by the severe facial expression. By learning a block weighting support vector machine, the component based approach is achieved. The proposed algorithm has two advantages: (1) only single one face training image is needed to train the classifier; (2) by using the facial block features with lower data dimensions, the proposed system is more computational efficiency. In particular, the proposed method achieves 97.94% face recognition accuracy when only using one training sample on the Yale B database. Experimental results demonstrate that the proposed method has reliable recognition rate when face images are under illumination and facial expression variations.

Road sign detection using eigen color

This paper presents a novel color-based method to detect road signs directly from videos. A road sign usually has specific colors and high contrast to its background. Traditional color-based approaches need to train different color detectors for detecting road signs if their colors are different. This paper presents a novel color model derived from Karhunen-Loeve (KL) transform to detect road sign color pixels from the background. The proposed color transform model is invariant to different perspective effects and occlusions. Furthermore, only one color model is needed to detect various road signs. After transformation into the proposed color space, a RBF (Radial Basis Function) network is trained for finding all possible road sign candidates. Then, a verification process is applied to these candidates according to their edge maps. Due to the filtering effect and discriminative ability of the proposed color model, different road signs can be very efficiently detected from videos. Experiment results have proved that the proposed method is robust, accurate, and powerful in road sign detection.

Improvement of Face Recognition by Eyeglass Removal

In this paper, we present a method based on Active Appearance Model (AAM) to remove eyeglasses from face images. The occluded regions are first roughly detected by the AAM search. Then, an ellipse model is used to fit the eyes’ position. After eliminating the eye blocks, residual regions are further enhanced and extracted more precisely by the adaptive threshold method. Finally, the synthesized image is generated by iterative error compensation. Experimental results show that the proposed method provides an effective solution to the recognition of faces occluded by thick-rimmed eyeglasses.

Suspicious object detection using fuzzy-color histogram

This paper proposes a novel method to detect suspicious objects from videos for abnormal event analysis. When considering a robbery event happens, there should be some suspicious object transferring conditions following between the forager and the victim. Since there is no prior knowledge about the objects property, it is difficult to automatically analyze the conditions without any manual efforts. To tackle this problem, a ratio histogram based on fuzzy c-means algorithm is proposed for finding suspicious objects. Furthermore, we use Gaussian mixture models to model the suspicious objects visual properties so that it can be accurately segmented from videos. After analyzing its subsequent motion features, different abnormal events like robbery can be effectively detected from videos. Experiment results have proved that the proposed method is robust, accurate, and powerful in abnormal event detection.

Human Body Part Segmentation of Interacting People by Learning Blob Models

In this paper, a scheme is proposed for solving segmentation problem when people engage in body contact in a video sequence. First, the body parts belonging to each interacting person are extracted using the deformable triangulation technique. The color blobs of each person are learned by Gaussian mixtures model on the fly before the person is interacting with another. Finally, those learned blob models are employed as decision criteria to segment each involved person out. The experimental results show that the proposed approach handles this kind of segmentation in an effective way.