• Proceedings of The KACE
• /
• 2017.08a
• /
• pp.209-211
• /
• 2017

기계학습은 인공지능(AI, Artificial Intelligence)의 일종으로 다른 인공지능 알고리즘이 정해진 규칙을 기반으로 주어진 임무(Task)를 해결하는 것과는 달리, 기계학습은 수집된 Data를 기반으로 최적의 솔루션을 학습한 후 미래의 값들을 예측하거나 해석하는 방법을 사용하고 있다. 더욱이 인터넷을 통한 연결성의 확대와 컴퓨터의 연산능력 발전으로 가능하게 된 Big-Data를 기반으로 하고 있어 이전의 인공지능 알고리즘에 비해 월등한 성능을 보여주고 있다. 그러나 기계학습 알고리즘이 Data를 학습할 때 학습 결과를 사람이 해석하기에 너무 복잡하여 사람이 그 내부 구조를 이해하는 것은 사실상 불가능하고, 이에 따라 학습된 기계학습 모델의 단점 또는 한계 등을 알지 못하는 문제가 있다. 본 연구에서는 이러한 블랙박스화된 기계학습 알고리즘의 특성을 이해하기 위해, 기계학습 알고리즘이 특정 입력에 대한 결과를 예측할 때 어떤 입력들로 부터 영향을 많이 받는지 그리고 어떤 입력으로부터 영향을 적게 받는지를 알아보는 방법을 소개하고 기존 연구의 단점을 개선하기 위한 방법을 제시한다.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.40 no.1
• /
• pp.152-164
• /
• 2015

A rapid development of semiconductors, sensors and mobile network technologies has enable that the embedded device includes high sensitivity sensors, wireless communication modules and a video processing module for vehicular environment, and many researchers have been actively studying the smart car technology combined on the high performance embedded devices. The vehicle is increased as the development of society, and the risk of accidents is increasing gradually. Thus, the advanced driver assistance system providing the vehicular status and the surrounding environment of the vehicle to the driver using various sensor data is actively studied. In this paper, we design and implement the smart vehicular camera device providing the V2X communication and gathering environment information. And we studied the method to create the metadata from a received video data and sensor data using video analysis algorithm. In addition, we invent S-ROI, D-ROI methods that set a region of interest in a video frame to improve calculation performance. We performed the performance evaluation for two ROI methods. As the result, we confirmed the video processing speed that S-ROI is 3.0 times and D-ROI is 4.8 times better than a full frame analysis.

• The Journal of The Korea Institute of Intelligent Transport Systems
• /
• v.7 no.3
• /
• pp.15-28
• /
• 2008

According to the accident statistics published by the National Police Agency in 2006, it can be recognized that drivers' characteristics and driving behaviors are the most causational factors on the traffic accidents. At present, although many recording tools such as digital speedometer or black box are distributed in the market to meet social requests of decreasing traffic accidents and increasing safe driving behaviors, it is also true that it still lacks in obvious categories for dangerous driving types and then, the efficiency of the categories to be studied has been low. In this study, dangerous driving types are redefined. They are grouped into 7 classifications in the first level, and the seven classifications are regrouped into 16 in more detail. To verify the redefined dangerous driving types, a Data-logger is developed to receive and analyze the data that occur from the driving behaviors of the test vehicle. The developed Data-logger can be used to construct a real time warning system and safe driving management system with dangerous driving patterns based on acceleration, deceleration, Yaw rate, image data, etc.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.20 no.9
• /
• pp.1778-1784
• /
• 2016

we improved the performance by parallelizing lane detection algorithms. Lane detection, as a intellectual assisting system, helps drivers make an alarm sound or revise the handle in response of lane departure. Four kinds of algorithms are implemented in order as following, Gaussian filtering algorithm so as to remove the interferences, gray conversion algorithm to simplify images, sobel edge detection algorithm to find out the regions of lanes, and hough transform algorithm to detect straight lines. Among parallelized methods, the data level parallelism algorithm is easy to design, yet still problem with the bottleneck. The high-speed data level parallelism is suggested to reduce this bottleneck, which resulted in noticeable performance improvement. In the result of applying actual road video of black-box on our parallel algorithm, the measurement, in the case of single-core, is approximately 30 Frames/sec. Furthermore, in the case of octa-core parallelism, the data level performance is approximately 100 Frames/sec and the highest performance comes close to 150 Frames/sec.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.10 no.3
• /
• pp.341-347
• /
• 2015

This study aims to design a chief VDR(Voyage Data Recorder) system which is appropriate to small and medium sized vessels and also implement the data about marine communication devices, sensors, etc. to be stored or printed at the navigator when those data are connected to VDR through data communication between marine navigation and VDR which are based on serial communication or internet in order to prove efficiency of the marine navigator. Also, the design of VDR is intended to be small and light in order to expand to apply it to small and medium vessels, which enables to analyze causes of marine accidents precisely through its characteristic functions which are the same as those at "vehicle mounted black-box" (location of the car, image and voice storage) by which the same roles are played on land.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.17 no.11
• /
• pp.2677-2685
• /
• 2013

Car running data collected from the vehicle is a native image data and sensing data of it. Hence, it can be used as a set of objective data based on which events that took place outside the car can be analyzed and determined. In this paper, we designed and implemented a emergency situation detection system to sense, store, and analyze signals related to car movements, driver's various operation states, collision pulse, etc when a car collision accident occurs on the actual road by sensing and analyzing the car movements and driver's operation status. The suggested system provides information on the driver's reaction right before the collision, operation state of the vehicle, and physical movement. The collected and analyzed data on vehicle running can be utilized to clarify the cause of a collision accident and to handle it in a just manner. Besides, it can contribute to grasping and correcting wrong driving habits of the driver and to saving.

• Journal of the Korea Computer Graphics Society
• /
• v.21 no.5
• /
• pp.11-19
• /
• 2015

Our research introduces a novel system for creating 3D vehicle animation. Our system is for intuitively authoring vehicle accident scenes according to videos or based on user-drawn trajectories. Our system has been implemented by combining three existing ideas. The first part is for obtaining 3D trajectory of a vehicle from black-box videos. The second part is a tracking algorithm that controls a vehicle to follow a given trajectory with small errors. The last part optimizes the vehicle control parameters so that the error between the input trajectory and simulated vehicle trajectory is minimized. We also simulate the deformation of the car due to an impact to achieve believable results in real-time.

• Journal of KIISE
• /
• v.43 no.1
• /
• pp.47-53
• /
• 2016

While the Internet develops rapidly, a huge amount of image data collected from smart phones, digital cameras and black boxes are being shared through social media sites. Generally, social images are handled by tagging them with information. Due to the ease of sharing multimedia and the explosive increase in the amount of tag information, it may be considered too much hassle by some users to put the tags on images. Image retrieval is likely to be less accurate when tags are absent or mislabeled. In this paper, we suggest a method of extracting tags from social images by using image content. In this method, CNN(Convolutional Neural Network) is trained using ImageNet images with labels in the training set, and it extracts labels from instagram images. We use the extracted labels for automatic image tagging. The experimental results show that the accuracy is higher than that of instagram retrievals.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.11 no.6
• /
• pp.664-670
• /
• 2018

In modern society, CCTV, which is the eye of surveillance, is being used to collect image data in various ways in daily life. CCTV is used not only for security, surveillance, and crime prevention but also in many fields such as automobile and black box. In this paper, we have developed a STM32F407 ARM chip based camera system for various applications. In order to develop camera system, modeling of camera system based on 3D structure was carried out in SolidWorks environment. The PCB board design was developed to extract the PCB parts from the camera system modeling files into iges files, convert them from the Altium Designer tool into 3D and 2D boards, After designing the camera system circuit and PCB, we have been studying the implementation of the stable system by using TRM (Thermal Risk Management) tool to cope with the heat simulation generated on the board.

• The Journal of Korea Institute of Information, Electronics, and Communication Technology
• /
• v.17 no.2
• /
• pp.113-120
• /
• 2024

The population sparseness index(PSI) is being utilized to describe the functioning of internal layers in artificial neural networks from the perspective of neurons, shedding light on the black-box nature of the network's internal operations. There is research indicating a positive correlation between the PSI and performance in each layer of convolutional neural network models for image classification. In this study, we observed the internal operations of a convolutional neural network when adversarial examples were applied. The results of the experiments revealed a similar pattern of positive correlation for adversarial examples, which were modified to maintain 5% accuracy compared to applying benign data. Thus, while there may be differences in each adversarial attack, the observed PSI for adversarial examples demonstrated consistent positive correlations with benign data across layers.