Youngtae Park

Vision-based remote control system by motion detection and open finger counting

In this paper, we present a universal remote control system based on computer vision. The method is composed of two stages of detecting visual evidences. Motion and skin color information is utilized to detect waving hands requesting control commands. Upon the request, the camera is controlled to zoom in on local region of the hand. The number of open fingers is counted by the shape analysis on the segmented hand region image. Control command is issued when a predetermined sequence of gesture state transition is produced. Experimental results show that the shape of open fingers exhibits strong features for determining correct gesture states. The use of stable features on consecutive frames yields robust and accurate performance regardless of operating conditions.

A comparison of neural net classifiers and linear tree classifiers: Their similarities and differences

Abstract Both neural net classifiers utilizing multilayer perceptron and linear tree classifiers composed of hierarchically structured linear discriminant functions can form arbitrarily complex decision boundaries in the feature space and have similar decision-making processes. The structure of the linear tree classifier can be easily mapped to that of the neural nets having two hidden layers by using the hyperplanes produced by the linear tree classifier. A new method for mapping the linear tree classifier to the neural nets having one hidden layer is presented with theoretical basis of mapping the convex decision regions produced by the linear tree classifier to the neurons in the neural nets. This mapping may be useful for choosing appropriately sized neural nets having one or two hidden layers, as well as for initializing the connection weights to speed up the learning rate and avoid the local trapping of the backpropagation algorithm. Also the internal operations of the hidden layer neurons in the three-layer feedforward nets could be well described by this mapping. Experimental results on both synthetic and real data suggest that this mapping is effective, and that there exists no significant difference in the classification accuracy of the neural net classifiers having one and two hidden layers. Substantial advantages of the neural net classifiers over the linear tree classifier were not found in the experiment, while the latter is much faster in both training and classifying stages.

Shape-resolving local thresholding for object detection

Abstract Selecting locally optimum thresholds, based on optimizing a criterion composed of the area variation rate and the compactness of the segmented shape is presented. The method is shown to have the shape-resolving property in the subtraction image, so that overlapped objects may be resolved into bright and dark evidences characterizing each object. As an application a vehicle detection algorithm robust to the operating conditions could be realized by applying simple merging rules to the geometrically correlated bright and dark evidences obtained by this local thresholding.

Measurement of traffic parameters in image sequence using spatio-temporal information

This paper proposes a novel method for measurement of traffic parameters, such as the number of passed vehicles, velocity and occupancy rate, by video image analysis. The method is based on a region classification followed by spatio-temporal image analysis. Local detection region images in traffic lanes are classified into one of four categories: the road, the vehicle, the reflection and the shadow, by using statistical and structural features. Misclassification at a frame is corrected by using temporally correlated features of vehicles in the spatio-temporal image. This capability of error correction results in the accurate estimation of traffic parameters even in high traffic congestion. Also headlight detection is employed for nighttime operation. Experimental results show that the accuracy is more than 94% in our test database of diverse operating conditions such as daytime, shadowy daytime, highway, urban way, rural way, rainy day, snowy day, dusk and nighttime. The average processing time is 30 ms per frame when four traffic lanes are processed, and real-time operation could be realized while ensuring robust detection performance even for high-speed vehicles up to 150 km h−1.

Discrete Hough transform using line segment representation for line detection

The Hough transform is a well-known method for detecting lines. The standard Hough transform (SHT) uses a straight line equation parameterized by an angle and a distance, so the axes of the Hough space are continuous and it is difficult to determine their optimal resolutions for digitization. To resolve this difficulty, we propose a discrete Hough transform (DHT) using line segment representation for line detection. The end points of line segments are used as parameters, so the resolutions of the axes can be easily determined and accumulated bins can be easily smoothed for noise suppression. Therefore, the proposed DHT is more appropriate for detecting isolated lines than the SHT.

A mapping from linear tree classifiers to neural net classifiers

Both neural net classifiers utilizing multilayer perceptron and linear tree classifiers composed of hierarchically structured linear discriminant functions can form arbitrarily complex decision boundaries in the feature space and have similar decision making process. The structure of the linear tree classifier can be easily mapped to that of the neural nets having two hidden layers by using the hyperplanes produced by the linear tree classifier. A new method for mapping the linear tree classifier to the neural nets having one hidden layer is presented with theoretical basis of mapping the convex decision regions produced by the linear tree classifier to the neurons in the neural nets. This mapping has been shown to be useful for choosing appropriately sized neural nets having one or two hidden layers.<<ETX>>

Vision-based object detection for passenger’s safety in railway platform

In this paper, we propose a vision-based object detection algorithm for railway passengerpsilas safety. The proposed algorithm uses three-dimensional position information with stereo cameras for minimizing various illuminant effects in railway platform environment, such as ambient illumination changes due to train arrival/departure in the scene. The detection process analyzes scene and detects both four different train status and fallen objects in preset monitoring area. To solve the detection problem in conventional two-dimensional detection system, the system detects object in three dimensionally by using stereo vision algorithm. We verify the system performance with extensive experimental results in a metro station. We expect the proposed algorithm will play a key role in establishing highly intelligent monitoring system for passengerpsilas safety for future railway environment.

3D Vision-based Security Monitoring for Railroad Stations

Increasing demands on the safety of public train services have led to the development of various types of security monitoring systems. Most of the surveillance systems are focused on the estimation of crowd level in the platform, thereby yielding too many false alarms. In this paper, we present a novel security monitoring system to detect critically dangerous situations such as when a passenger falls from the station platform, or when a passenger walks on the rail tracks. The method is composed of two stages of detecting dangerous situations. Objects falling over to the dangerous zone are detected by motion tracking. 3D depth information retrieved by the stereo vision is used to confirm fallen events. Experimental results show that virtually no error of either false positive or false negative is found while providing highly reliable detection performance. Since stereo matching is performed on a local image only when potentially dangerous situations are found; real-time operation is feasible without using dedicated hardware.

Robust vehicle detection based on shadow classification

The multi-level shadow classification has been shown to provide reliable information on the presence of vehicles in traffic scenes. The method is based on classifying the shadow shapes into six categories at each threshold level. Non-overlapping shadow shapes with higher priority are selected at each level. Shadow-reshaping capability makes the resulting shadow information robust to the variation of operating conditions. Unlike other approaches, vehicle movement information between frames is not utilized; thereby the traffic parameters can be measured quantitatively even when the vehicle movement is not observed. Also the detecting performance is not affected by the abrupt change of weather because background information is not utilized

An ART2 trained by two-stage learning on circularly ordered data sequence

Adaptive Resonance Theory (ART), characterized by its built-in mechanism of handling the stability-plasticity dilemma and by fast adaptive learning without forgetting informations learned in the past, is based on an unsupervised template matching. We propose an improved two-stage learning algorithm for ART2: the original unsupervised learning followed by a new supervised learning. Each of the output nodes, after the unsupervised learning, is labeled according to the category informations of the feature vectors associated with the node. In the supervised learning, each feature vector is used to reinforce the template pattern associated with the target output node belonging to the same category as the feature vector. Another modification is a circular ordering of the training sequence, which is intended to prevent some dominant classes from exhausting a finite number of template patterns in ART2. The proposed learning algorithm has been shown to yield better accuracy than the original ART2, regardless of the size of the network. The hold-out recognition accuracy of the modified ART2 on the real data obtained from military ship images is 98.4%, and that of the original ART is 94.8%, when the size of the network is chosen reasonably in such a way that the size is minimized while maintaining the required accuracy.<<ETX>>