While human activity recognition and pose estimation are closely related, these two issues are usually treated as separate tasks. In this thesis, two-dimension and three-dimension pose estimation is obtained for human activity recognition in a video sequence, and final activity is determined by combining it with an activity algorithm with visual attention. Two problems can be solved efficiently using a single architecture. It is also shown that end-to-end optimization leads to much higher accuracy than separated learning. The proposed architecture can be trained seamlessly with different categories of data. For visual attention, soft visual attention is used, and a multilayer recurrent neural network using long short term memory that can be used both temporally and spatially is used. The image, pose estimated skeleton, and RGB-based activity recognition data are all synthesized to determine the final activity to increase reliability. Visual attention evaluates the model in UCF-11 (Youtube Action), HMDB-51 and Hollywood2 data sets, and analyzes how to focus according to the scene and task the model is performing. Pose estimation and activity recognition are tested and analyzed on MPII, Human3.6M, Penn Action and NTU data sets. Test results are Penn Action 98.9%, NTU 87.9%, and NW-UCLA 88.6%.

Understanding occupant behavior is critical for enhancing safety and situational awareness in intelligent transportation systems. This study investigates multimodal occupant behavior recognition using sequential inputs extracted from 2D pose, 2D gaze, and facial movements. We conduct a comprehensive comparative study of three distinct architectural paradigms: a static Multi-Layer Perceptron (MLP), a recurrent Long Short-Term Memory (LSTM) network, and an attention-based Transformer encoder. All experiments are performed on the large-scale Occupant Behavior Classification (OBC) dataset, which contains approximately 2.1 million frames across 79 behavior classes collected in a controlled, simulated environment. Our results demonstrate that temporal models significantly outperform the static baseline. The Transformer model, in particular, emerges as the superior architecture, achieving a state-of-the-art Macro F1 score of 0.9570 with a configuration of a 50-frame span and a step size of 10. Furthermore, our analysis reveals that the Transformer provides an excellent balance between high performance and computational efficiency. These findings demonstrate the superiority of attention-based temporal modeling with multimodal fusion and provide a practical framework for developing robust and efficient in-vehicle occupant monitoring systems. Implementation code and supplementary resources are available (see Data Availability Statement).

Image classification is a fundamental task in computer vision, with convolutional neural networks (CNNs) being the state-of-the-art approach for this task. However, CNNs can be large and computationally expensive, which limits their deployment on resource-constrained devices. To address this issue, model optimization and compression techniques such as knowledge distillation and channel attention have been proposed. In this paper, we propose a modified version of the skip-connection knowledge distillation method, which integrates the efficient channel attention module. Our experiments on the CIFAR-10 dataset show that this approach improves the performance of both large teacher-small student and self-distillation scenarios, and also leads to smaller model sizes. We also visualize the intermediate feature maps to understand the learning process of the network. Our results demonstrate the effectiveness of combining knowledge distillation and channel attention for improving the performance and efficiency of CNNs.