Chanjun Chun

Accurately identifying whether a vehicle is controlled by a human driver or by an autonomous policy is critical in future mixed-traffic environments for safety monitoring and liability attribution. To achieve accurate classification or identification, this paper proposes NMTT-Net (Noise-agnostic Multimodal Temporal Transformer Network), a multimodal time-series classifier that combines VAE-derived latent embeddings from front-view camera images with 18-dimensional vehicle dynamics signals. To maintain reliability under adverse weather and sensor corruption, NMTT-Net applies modality-specific noise detection and denoising prior to temporal encoding, and learns discriminative behavioral patterns using dual-path Temporal Transformer encoders with fixed-weight fusion. We construct a Human-to-Autonomy (H2A) dataset in the CARLA simulator under Town07 containing Human, PPO, and DDQN trajectories under clean and noisy conditions, in which image noise is induced by adverse weather and dynamics noise is modeled as Gaussian corruption. Experimental results show that NMTT-Net achieves 99.509% and 98.617% accuracy under clean and noisy conditions, respectively, resulting in a robustness gap (clean-to-noisy accuracy drop) of 0.892%, which is smaller than that of denoised single-modality baselines. These simulation-bound results demonstrate that noise-aware multimodal temporal fusion can enable reliable H2A identification under severe environmental degradation, although further study is necessary for real-world readiness.

Passive SONAR has attracted significant attention in covert operations, enabling vessel detection and classification through spectrogram analysis of received signals without active transmission. However, in realistic ocean environments, interference from multiple vessels and ambient noise make accurate analysis highly challenging. As deep learning-based source separation models rely on time–frequency representations, they often fail in complex environments, making spatial cues crucial for accurate target separation. Therefore, we propose BTSTNet, a beamforming-based transformer network that performs spatially aware target separation from complex acoustic mixtures. BTSTNet processes multi-channel beamformed inputs as a joint space–time–frequency representation and separates the target signal from complex mixtures in the target direction. We integrate a hybrid transformer encoder–decoder architecture with a novel Cross-Channel Aggregation Head (CCA-Head) to model spatial dependencies across channels. In addition, we construct a new beamforming-based passive SONAR mixture dataset that simulates diverse underwater conditions. Experimental evaluation on the constructed simulation dataset indicates that BTSTNet achieves superior separation performance compared to existing deep learning-based source separation models.

Underwater acoustic signal denoising technology aims to overcome the challenge of recovering tonal patterns of target sources such as ships and submarines in complex ocean environments. Conventional statistical methods and deep learning approaches suffer from significant performance degradation under low signal-to-noise ratio (SNR) conditions, and their practical application is further limited by the scarcity of labeled training data. In this paper, we propose an adversarial representation learning framework based on the LOFARgram representation to suppress underwater acoustic noise while preserving tonal structures. To tackle the limited training data problem, we generate simulation data by modeling vessel acoustics and environmental noise based on Wenz curves, augmented with the ShipsEar dataset to improve generalization to real-world conditions. For tonal pattern preservation, we adopt an adversarial learning scheme with an metric-based discriminator to explicitly maintain structural consistency in reconstructed LOFARgrams. To validate the robustness of the proposed method, experiments are conducted on U-Net variants and SE-Mamba architectures with distinct structural properties. Experimental results demonstrate that the proposed method substantially outperforms baseline configurations across various architectures. Qualitative evaluation on real-world recordings suggests practical applicability to operational underwater acoustic denoising.

Reliable detection of unknown signals is essential for ensuring the robustness of underwater acoustic sensing systems, particularly in maritime security and autonomous navigation. However, Conventional deep learning models often exhibit overconfidence when encountering unknown signals and are unable to quantify predictive uncertainty due to their deterministic inference process. To address these limitations, this study proposes a novelty detection framework that integrates an out-of-distribution detector for neural networks (ODIN) with Monte Carlo (MC) dropout. ODIN mitigates model overconfidence and enhances the separability between known and unknown signals through softmax probability calibration, while MC dropout introduces stochasticity via multiple forward passes to estimate predictive uncertainty—an element critical for stable sensing in real-world underwater environments. The resulting probabilistic outputs are modeled using Gaussian mixture models fitted to ODIN-calibrated softmax distributions of known classes. The Kullback–Leibler divergence is then employed to quantify deviations of test samples from known class behavior. Experimental evaluations on the DeepShip dataset demonstrate that the proposed method achieves, on average, a 9.5% and 5.39% increase in area under the receiver operating characteristic curve, and a 7.82% and 2.63% reduction in false positive rate at 95% true positive rate, compared to the MC dropout and ODIN baseline, respectively. These results confirm that integrating stochastic inference with ODIN significantly enhances the stability and reliability of novelty detection in underwater acoustic environments.

Egocentric hand gesture recognition is vital for natural human–computer interaction in augmented and virtual reality (AR/VR) systems. However, most deep learning models struggle to balance accuracy and efficiency, limiting real-time use on wearable devices. This paper introduces a Two-Stream Temporal Shift Module Transformer Fusion Network (TSMTFN) that achieves high recognition accuracy with low computational cost. The model integrates Temporal Shift Modules (TSMs) for efficient motion modeling and a Transformer-based fusion mechanism for long-range temporal understanding, operating on dual RGB-D streams to capture complementary visual and depth cues. Training stability and generalization are enhanced through full-layer training from epoch 1 and MixUp/CutMix augmentations. Evaluated on the EgoGesture dataset, TSMTFN attained 96.18% top-1 accuracy and 99.61% top-5 accuracy on the independent test set with only 16 GFLOPs and 21.3M parameters, offering a 2.4–4.7× reduction in computation compared to recent state-of-the-art methods. The model runs at 15.10 samples/s, achieving real-time performance. The results demonstrate robust recognition across over 95% of gesture classes and minimal inter-class confusion, establishing TSMTFN as an efficient, accurate, and deployable solution for next-generation wearable AR/VR gesture interfaces.

Fifth-generation (5G) mobile network is intended to facilitate high data rates with massive connectivity with the benefit of small cell technology. The cloud radio access network (C-RAN) is a promising mobile network architecture that can meet the ever-increasing resource demand of growing number of users. In C-RAN, the base station functionality is split into a baseband unit (BBU) and remote radio head (RRH); BBUs from multiple sites are centralized and virtualized utilizing cloud computing and virtualization techniques. However, resource allocation to the huge amount of dynamic users is a tremendous challenge to reach the objectives of system efficiency and meeting dynamic user requirements. In this paper, we present a scheme of allocating resources in C-RAN that combines centralized and multi-agent-based resource allocation. The centralized controller monitors and operates from inside the BBU pool, while the virtual base stations (VBSs) connected to the distributed RRHs act as agents in the multi-agent system (MAS). The ultimate resource allocation solution is derived from balancing between resources requested by MAS agents and resource demand approximation from historical data by the centralized controller. We conduct a simulation-based study to evaluate the performance of our proposed mechanism and compare with other resource allocation schemes.

The next generation of mobile networks is predicted to deliver high data speeds, lower latency, and increase the spectral and energy efficiency of wireless communication systems. Several technologies are being investigated for usage in 5G networks. Massive multiple-input multiple-output (mMIMO) systems are one of the most promising technologies for enabling 5G. Even after the recent advancements and research, numerous challenges still exist for channel estimation for mMIMO systems. In the context of pilot contamination and feedback overhead, this study tracks the most recent developments in research on mMIMO system difficulties. The primary goals of this study are to identify the problems with channel estimation, provide a summary of the cutting-edge solutions suggested in the literature, and then discuss newly emerging open research issues that must be taken into account for the implementation of beyond-5G networks.

Speaker recognition is a technology that identifies the speaker in an input utterance by extracting speaker-distinguishable features from the speech signal. Speaker recognition is used for system security and authentication; therefore, it is crucial to extract unique features of the speaker to achieve high recognition rates. Representative methods for extracting these features include a classification approach, or utilizing contrastive learning to learn the speaker relationship between representations and then using embeddings extracted from a specific layer of the model. This paper introduces a framework for developing robust speaker recognition models through contrastive learning. This approach aims to minimize the similarity to hard negative samples—those that are genuine negatives, but have extremely similar features to the positives, leading to potential mistaken. Specifically, our proposed method trains the model by estimating hard negative samples within a mini-batch during contrastive learning, and then utilizes a cross-attention mechanism to determine speaker agreement for pairs of utterances. To demonstrate the effectiveness of our proposed method, we compared the performance of a deep learning model trained with a conventional loss function utilized in speaker recognition with that of a deep learning model trained using our proposed method, as measured by the equal error rate (EER), an objective performance metric. Our results indicate that when trained with the voxceleb2 dataset, the proposed method achieved an EER of 0.98% on the voxceleb1-E dataset and 1.84% on the voxceleb1-H dataset.

In this paper, we propose a method for estimating the classes and directions of static audio objects using stereo microphones in a drone environment. Drones are being increasingly used across various fields, with the integration of sensors such as cameras and microphones, broadening their scope of application. Therefore, we suggest a method that attaches stereo microphones to drones for the detection and direction estimation of specific emergency monitoring. Specifically, the proposed neural network is configured to estimate fixed-size audio predictions and employs bipartite matching loss for comparison with actual audio objects. To train the proposed network structure, we built an audio dataset related to speech and drones in an outdoor environment. The proposed technique for identifying and localizing sound events, based on the bipartite matching loss we proposed, works better than those of the other teams in our group.