We propose a vision-based fall detection algorithm using advanced deep learning models and fusion methods for smart safety management systems. By detecting falls through visual cues, it is possible to leverage existing surveillance cameras, thus minimizing the need for extensive additional equipment. Consequently, we developed a cost-effective fall detection system. The proposed system consists of four modules: object detection, pose estimation, action recognition, and result fusion. Constructing the fall detection system involved the utilization of state-of-the-art (SOTA) models. In the fusion module, we experimented with various approaches, including voting, maximum, averaging, and probabilistic fusion. Notably, we observed a significant performance improvement with the use of probabilistic fusion. We employed the HAR-UP dataset to demonstrate this enhancement, achieving an average 0.84% increase in accuracy compared to the baseline, which did not incorporate fusion methods. By applying our proposed time-level ensemble and skeleton-based fall detection approach, coupled with the use of enhanced object detection and pose estimation modules, we substantially improved the robustness and accuracy of the system, particularly for fall detection in challenging scenarios.

Gait recognition is a biometric technology that identifies individuals based on the unique characteristics of their gait. With the advancement of deep learning based computer vision technology, gait recognition has significantly improved in performance and gained significant attention due to its non-invasive nature. This paper proposes a vision-based gait recognition model that enhances performance by applying deformable convolution in the temporal dimension. Gait is a periodic behavior, and understanding the body’s changing patterns during walking is crucial for improving recognition accuracy. Important features for gait recognition can appear in both short and long moments of the walking process. Therefore, this study proposes a model that performs deformable convolution in the temporal dimension to effectively extract such features. This allows for a more flexible understanding of temporal patterns, eliminating the need for hand-crafted design of filters for extracting temporal features and enabling the model to learn important features for gait recognition that appear in both short and long time intervals in a unified manner. As a result, the proposed model achieved state-of-the-art performance on the CASIA-B dataset.

The objective of this work is to address the problem of detecting track intruders in railway systems using deep learning-based algorithms. Unauthorized entry onto railway tracks poses a significant risk of collisions between trains and humans. However, intrusion discrimination algorithms often suffer from a lack of learning data and data imbalance issues. To overcome these challenges, this research proposes an algorithm that combines generative models and classification networks. Generative models are utilized to generate synthetic intrusion data by learning the underlying distribution of available data and creating new samples resembling the original data. The augmented intrusion data is then used to train deep neural networks to accurately identify intrusions. The proposed algorithm is evaluated using real data sets, demonstrating its effectiveness in overcoming limited learning data and data imbalance issues. By augmenting intrusion data using generative models, the algorithm achieves improved accuracy compared to traditional approaches. In conclusion, the algorithm presented in this work provides a solution for detecting track intruders in railway systems. By leveraging generative models to augment limited intrusion data and utilizing classification networks for intrusion discrimination, the algorithm demonstrates improved performance in accurately identifying intrusions. This research highlights the potential of deep learning-based approaches in enhancing railway safety and recommends further exploration and application of these methods in real-world settings.

Laser welding plays a critical role in advanced manufacturing, particularly in battery pack assembly for electric vehicles, where weld quality directly impacts performance and safety. However, current inspection methods rely on manual visual, electrical, and physical checks, resulting in inconsistent outcomes and limited detection of latent defects undetectable by traditional techniques. This study proposed a two-stage deep learning-based quality inspection framework designed to remain effective even when defect data were scarce. In the first stage, a convolutional neural network-based autoencoder was trained exclusively on normal data and augmented with a novel loss function that captured distributional features. In the second stage, multi-class defect classification was performed using the encoder and bottleneck layers of the first stage as a shared backbone, enabling robust performance in data-constrained environments. Data augmentation techniques were further applied to improve the generalization capability. A high-quality dataset collected from an actual industrial setting was used to validate the proposed approach. Experimental results showed that the proposed two-stage framework achieved 100% accuracy and an F1-score of 1.0 in defect detection and classification, outperforming the conventional rule-based system (99.31% accuracy and 0.995 F1-score) and demonstrating perfect consistency across all test samples. The proposed framework offers a scalable and intelligent solution for automated inspection in industrial laser welding applications.

In recent years, the application of pretrained models in specialized domains has become increasingly important. Traditionally, adapting these models involved fine-tuning their parameters and structures through retraining. However, these fine-tuning methods can be inefficient, particularly when addressing data from specific domains or when modifications are needed in the lower layers of large-scale pretrained models. This study aims to investigate the effectiveness of using pretrained models with frozen weights for downstream tasks in the context of railway track detection, particularly focusing on the railway system. To achieve this, we employed a large-scale semantic segmentation model that had been pretrained on extensive datasets. The models utilized were kept with fixed weights, eliminating the need for retraining. We conducted a comparative analysis of various pretrained models sourced from different datasets to identify the most suitable model for the track detection system. The findings from our experiments revealed the performance metrics of the selected pretrained models, highlighting their effectiveness in the specific domain of railway track detection. Overall, this research demonstrates the practical applicability of pretrained models with frozen weights in specialized fields such as railway systems, offering insights into their usefulness and potential for improving detection algorithms in this domain.