Search

This paper tackles the problem of predicting 3D Human Pose on AI Edge Devices. An AI edge device is an AI-enabled Internet of Things (IoT) device that can run Deep Learning (DL) models right on the device without connecting to the Internet. In recent years, researchers have proposed numerous DL-based models for 3D human pose estimation (HPE), but no work focuses on solving this task in edge devices. Building a DL model for an edge device has unique challenges such as limited computation capacity, small memory, and low power. This paper investigates how to optimize a big, heavy-computing 3D pose estimation DL model into a light-weight, small-size model that can run efficiently in a device. Specially, we propose an End-to-End pipeline to run 3D human pose prediction (HPP) in Real-time on AI edge devices. Furthermore, our proposed end-to-end pipeline is general and can be employed by other AI edge device based real-world applications.

The assume-guarantee verification has been recognized as a promising method for solving the state space explosion in modular model checking of component-based software. However, the counterexample analysis technique used in this method has huge complexity and the computational cost for generating assumptions is very high. As a result, the method is difficult to be applied in practice. Therefore, this paper presents two improvements of the assume-guarantee verification method in order to solve the above problems. The first one is a counterexample analysis method that is simple to implement but effective enough to prevent the verification process from infinite loops when considering the last action of counterexample as suffix in implementation. This is done by finding a suffix that can make the observation table not closed when being added to the suffix set of the table and use that suffix for the learning process. The second one is a reduction of the number of membership queries to be asked to teacher when learning assumptions. This results in a significantly faster speed in generating assumption than that of the original algorithm. An implemented tool and experimental results are also described to show the effectiveness of the improvements.

This paper proposes an efficient method for automated generating models of component-based software. This method accepts regular expressions that describe behaviors of software components. The proposed method uses the CNNFA algorithm to parse the regular expressions in order to generate corresponding models. This method can generate minimal accurate models of the software components. A tool is implemented and applied for some typical systems to show the efficiency of this method. The obtained experimental results show that this method is faster than existing methods. The generated models play an important role in making such model-based approaches as model checking and model-based testing more feasible in practice.