publications

2022

1. Sensors

   Moving object tracking based on sparse optical flow with moving window and target estimator

   Hosik Choi, Byungmun Kang, and DaeEun Kim

   Sensors, 2022

   Abs DOI HTML PDF

   Moving object detection and tracking are technologies applied to wide research fields including traffic monitoring and recognition of workers in surrounding heavy equipment environments. However, the conventional moving object detection methods have faced many problems such as much computing time, image noises, and disappearance of targets due to obstacles. In this paper, we introduce a new moving object detection and tracking algorithm based on the sparse optical flow for reducing computing time, removing noises and estimating the target efficiently. The developed algorithm maintains a variety of corner features with refreshed corner features, and the moving window detector is proposed to determine the feature points for tracking, based on the location history of the points. The performance of detecting moving objects is greatly improved through the moving window detector and the continuous target estimation. The memory-based estimator provides the capability to recall the location of corner features for a period of time, and it has an effect of tracking targets obscured by obstacles. The suggested approach was applied to real environments including various illumination (indoor and outdoor) conditions, a number of moving objects and obstacles, and the performance was evaluated on an embedded board (Raspberry pi4). The experimental results show that the proposed method maintains a high FPS (frame per seconds) and improves the accuracy performance, compared with the conventional optical flow methods and vision approaches such as Haar-like and Hog methods.

2. ICEE

   Efficient Data Selection with YOLOv5 Deep Learning Model

   Hosik Choi, Byungmun Kang, and DaeEun Kim

   2022

   Abs PDF

   The state of the art deep learning models such as YOLOv5 have been applied to a wide research field including object detection. The YOLOv5 trains a relatively large size of neural networks and it needs much training time and lots of datasets. In this paper, we propose a strategy of efficient data selection to reduce efforts of data acquisition. The data selection method can be a useful tool for deep learning on a large amount of datasets. If there is no information of data characteristics for training data, all types of datasets should be acquired to train neural networks. We tested YOLOv5 deep learning for a variety of datasets including six different types of custom image datasets, and investigated how a portion of image dataset for a given type can affect the performance of object detection for another type. From the results, we can choose a proper set of training data with a relatively small loss of performance.

3. Master’s Thesis

   An Integrative Approach to Pedestrian Detection Based on Sparse Optical Flow and Convolutional Neural Network Model

   Hosik Choi

   2022

   Abs PDF

   In this dissertation, a light-weight object detection algorithm is designed to perform efficiently, even in situations where computer performance is limited.