Abstract:
"Over the last few years, the realism of graphics in the 3D entertainment industry has been tremendously increasing due to the remarkable advancements in relation to artistic techniques and artificial intelligence. One of the major considerations of the characters in virtual worlds is how the depth of the character’s feelings in a specific environment should be expressed, as it needs to connect with the player or user’s cognitive experiences mutually. A significant problem will arise on how the massive worlds should play the role of influencers to visualize interconnection between the virtual character and the world.
The key features to identify those aspects are the movements of the virtual character and dynamic changes of the virtual environment variations. To mimic the real-life abstraction into a virtual world, a novel approach has been proposed to build the relationship utilizing psychological frameworks and interactive techniques. Multiple data-driven approaches have been experimented with to satisfy those requirements by utilizing the benefits of Modern deep learning techniques and their advancement in this field of study. Ultimately, traditional classification and regression methodologies were utilized in a way that performs multiple tasks in a single model using different modalities at the same time as the design has been finalized considering the core factors of the PAD emotional model, locomotion states and the visual attributes of the virtual world.
The author has successfully produced a robust model to predict the current surrounding environment type and its mood factor values in a real-time 3D game engine using the architecture built by Convolutional and Multi-Layer Perceptron (MLP) networks. The classification of the environment type and accuracy held up to 98%, and the Pleasure, Arousal and Dominance deviated within the range of a maximum 10 - 17 % error rate of MAE, MSE and RMSE.
"
