Abstract This paper aims to describe a bootloader for embedded devices. The purpose of the bootloader is to be the first to run when the microcontroller starts and to provide a way of controlling what application runs next, similarly to how bootloaders work on the desktop systems. The difference here being that the bootloader’s purpose is to allow the embedded device... This paper aims to describe a bootloader for embedded devices. The purpose of the bootloader is to be the first to run when the microcontroller starts and to provide a way of controlling what application runs next, similarly to how bootloaders work on the desktop systems. The difference here being that the bootloader’s purpose is to allow the embedded device to be much more flexible, since the application running on the microcontroller can be changed without using specialized hardware. This flexibility and the fact that no specialized hardware programmers are needed also leads to a reduction of costs for deploying and maintaining the embedded systems. These two goals, flexibility and cost reduction must be accompanied by a third one: security. The recent growth of embedded systems in the IoT (Internet of Things) domain demands that the device deployed be more secure than in the past. Read More Read Less

Abstract To increase the performance of sport team, the tactics analysis of team from game video is essential. Trajectories of the players are the most useful cues in a sport video for tactics analysis. In this paper, we propose a technique to reconstruct the trajectories of players from broadcast basketball videos. We first propose a mosaic based approach to detect the boundary lines of... To increase the performance of sport team, the tactics analysis of team from game video is essential. Trajectories of the players are the most useful cues in a sport video for tactics analysis. In this paper, we propose a technique to reconstruct the trajectories of players from broadcast basketball videos. We first propose a mosaic based approach to detect the boundary lines of court. Then, the locations of players are determined by the integration of shape and color visual information. A layered graph is constructed for the detected players, which includes all possible trajectories. A dynamic programming based algorithm is applied to find the trajectory of each player. Finally, the trajectories of players are displayed on a standard basketball court model by a homography transformation. In contrast to related works, our approach exploits more spatiotemporal information in video. Experimental results show that the proposed approach works well and outperforms some existing technique. Read More Read Less

Abstract Linear models with nonlinear attribute modeling are popularly selected for massive regression and categorization tasks with sparse inputs. Memorization of attribute interactions through a broad batch of cross-product attribute conversion is compelling and explainable, while generalization needs extra feature engineering exercise. With limited effort, deep neural networks (DNN) can be generalized exceptionally to undiscovered feature sequences. However, DNN can over-generalize and recommend lesser suited... Linear models with nonlinear attribute modeling are popularly selected for massive regression and categorization tasks with sparse inputs. Memorization of attribute interactions through a broad batch of cross-product attribute conversion is compelling and explainable, while generalization needs extra feature engineering exercise. With limited effort, deep neural networks (DNN) can be generalized exceptionally to undiscovered feature sequences. However, DNN can over-generalize and recommend lesser suited items. In this paper, we propose an ensemble of various collectively trained linear models and DNNs to blend the advantage of memorization and generalization for a recommender system (RS). We evaluated the system on MovieLens 100K Dataset, a stable benchmark dataset with 100,000 ratings on 1682 movies from 943 users. Our experiment results show that ensembling of wide & deep models significantly increased recommendation accuracy and decreased mean absolute error (MAE) and root mean square error (RMSE) in comparison with collaborative filtering techniques, deep-only and wide-only models. Read More Read Less