In this paper, we implemented a deep learning-based automatic object tracking and handy motion control drone system and analyzed the performance of the proposed system. The drone system automatically detects and tracks targets by analyzing images obtained from the drone's camera using deep learning algorithms, consisting of the YOLO, the MobileNet, and the deepSORT. Such deep learning-based detection and tracking algorithms have both higher target detection accuracy and processing speed than the conventional color-based algorithm, the CAMShift. In addition, in order to facilitate the drone control by hand from the ground control station, we classified handy motions and generated flight control commands through motion recognition using the YOLO algorithm. It was confirmed that such a deep learning-based target tracking and drone handy motion control system stably track the target and can easily control the drone.

It is difficult to accurately extract features of pedestrian because the pedestrian is photographed at a crosswalk using a camera positioned higher than the pedestrian. In addition, it is more difficult to extract features when a part of the pedestrian's body is covered by an umbrella or parasol or when the pedestrian is holding an object. Representative methods to solve this problem include Object Detection, Instance Segmentation, and Pose Estimation. Among them, this study intends to use the Pose Estimation method. In particular, we intend to increase the recognition rate of pedestrians in crosswalks by maintaining the image resolution through HigherHRNet and applying the foot estimation technique. Finally, we show the superiority of the proposed method by applying and analyzing several data sets covered by body parts to the existing method and the proposed method.

For stereo matching based on deep learning, the design of network structure is crucial to the calculation of matching cost, and the time-consuming problem of convolutional neural network in image processing also needs to be solved urgently. In this paper, a method of stereo matching using sparse loss volume in parallax dimension is proposed. A sparse 3D loss volume is constructed by using a wide step length translation of the right view feature map, which reduces the video memory and computing resources required by the 3D convolution module by several times. In order to improve the accuracy of the algorithm, the nonlinear up-sampling of the matching loss in the parallax dimension is carried out by using the method of multi-category output, and the training model is combined with two kinds of loss functions. Compared with the benchmark algorithm, the proposed algorithm not only improves the accuracy but also shortens the running time by about 30%.

Our final goal is to implement nursing robots that can recognize patient's faces and their medicine on prescription. They can help patients to take medicine on time and prevent its abuse for recovering their health soon. As the first step, we proposed a medicine classifier with a low computational network that is able to run on embedded PCs without GPU in order to be applied to universal nursing robots. We confirm that our proposed model called MedicineNet achieves an 99.99% accuracy performance for classifying 15 kinds of medicines and background images. Moreover, we realize that the calculation time of our MedicineNet is about 8 times faster than EfficientNet-B0 which is well known as ImageNet classification with the high performance and the best computational efficiency.

In this paper, we propose a method for a mobile robot to estimate a specific location of a service provision target using a beacon-tag for the purpose of providing location-based services (LBS) to users in an indoor environment. To estimate the location, the irregular characteristics and error factors of the received signal strength indicator (RSSI) generated from the beacon are analyzed, and the distance conversion function is derived from the RSSI data extracted by applying a Gaussian filter. Then, the distance data converted from the plurality of beacons estimates an indoor location through a triangulation technique. After that, the improvement in the location estimation is analyzed by applying the temporal confidence reasoning technique. The possibility of providing a LBS of a mobile robot was confirmed through a location estimation experiment for a plurality of designated locations in an indoor environment.

Edge Computing enables image-based inference in close proximity to end users and real-world objects. However, since edge servers have limited computational and accelerator resources, efficient resource management is essential. In this paper, we present HeteroAccel system that performs optimal scheduling in Kubernetes platform based on available node and accelerator information for various inference requests. Our experiments showed 25.3% improvement in overall inference performance over the default scheduling scheme in edge computing environment in which four types of inference services are requested.

Panoptic segmentation, which is now widely used in computer vision such as medical image analysis, and autonomous driving, helps understanding an image with holistic view. It identifies each pixel by assigning a unique class ID, and an instance ID. Specifically, it can classify 'thing' from 'stuff', and provide pixel-wise results of semantic prediction and object detection. As a result, it can solve both semantic segmentation and instance segmentation tasks through a unified single model, producing two different contexts for two segmentation tasks. Semantic segmentation task focuses on how to obtain multi-scale features from large receptive field, without losing low-level features. On the other hand, instance segmentation task focuses on how to separate 'thing' from 'stuff' and how to produce the representation of detected objects. With the advances of both segmentation techniques, several panoptic segmentation models have been proposed. Many researchers try to solve discrepancy problems between results of two segmentation branches that can be caused on the boundary of the object. In this survey paper, we will introduce the concept of panoptic segmentation, categorize the existing method into two representative methods and explain how it is operated on two methods: top-down method and bottom-up method. Then, we will analyze the performance of various methods with experimental results.

3D data can be categorized into two parts : Euclidean data and non-Euclidean data. In general, 3D data exists in the form of non-Euclidean data. Due to irregularities in non-Euclidean data such as mesh and point cloud, early 3D deep learning studies transformed these data into regular forms of Euclidean data to utilize them. This approach, however, cannot use memory efficiently and causes loses of essential information on objects. Thus, various approaches that can directly apply deep learning architecture to non-Euclidean 3D data have emerged. In this survey, we introduce various deep learning methods for mesh and point cloud data. After analyzing the operating principles of these methods designed for irregular data, we compare the performance of existing methods for shape classification and segmentation tasks.

Deraining is one of the image restoration tasks and should consider a tradeoff between local details and broad contextual information while recovering images. Current studies adopt an attention mechanism which has been actively researched in natural language processing to deal with both global and local features. This paper classifies existing deraining methods and provides comparative analysis and performance comparison by using several datasets in terms of generalization.

We propose a novel, highly accurate approximate multiplier using different types of inexact 4-2 compressors. The importance of low hardware costs leads us to develop approximate multiplication for error-resilient applications. Several rules are developed when selecting a topology for designing the proposed multiplier. Our highly accurate multiplier design considers the different error characteristics of adopted compressors, which achieves a good error distribution, including a low relative error of 0.02% in the 8-bit multiplication. Our analysis shows that the proposed multiplier significantly reduces power consumption and area by 45% and 26%, compared with the exact multiplier. Notably, a trade-off relationship between error characteristics and hardware costs can be achieved when considering those of existing highly accurate approximate multipliers. In the image blending, edge detection and image sharpening applications, the proposed 8-bit approximate multiplier shows better performance in terms of image quality metrics compared with other highly accurate approximate multipliers.