Egocentric Vision Action Recognition: Performance Analysis on the Coer_Egovision Dataset

Egocentric activity recognition, also known as first-person vision, captures human actions and activities from the perspective of a wearable camera, providing a personalized and contextualized viewpoint. This unique perspective is crucial for various real-world applications, including healthcare monitoring, sports analy-sis, augmented reality, and assistive technologies. In this paper, we present a comparative analysis of various state-of-the-art algorithms for egocentric action recognition using the recently curated dataset, “Coer_Egovision.” This dataset captures human actions and activities from the unique perspective of a wearable camera, providing a personalized and context-aware viewpoint. Such a perspective holds significant value for real-world applications, including healthcare monitoring, sports analysis, augmented reality, and assistive technologies.The “Coer_Egovision” dataset encompasses diverse actions, such as “Going_Downstairs”, “Going_Upstairs”, “Texting”, “Walking” and “Writing” allowing for a challenging benchmark to evaluate algorithms in this domain. By leveraging the wearable camera perspective, the dataset addresses the need for a more com-prehensive understanding of human interactions with the environment.Among the evaluated algorithms, the Long-term Recurrent Convolutional Network (LRCN) emerges as the top performer, achieving an impressive accuracy of 88% in effectively modeling long-term temporal dependencies. Additionally, the Histogram of Oriented Gradients (HOG) combined with Support Vector Machine (SVM) approach demonstrates robust performance with 82% accuracy, highlighting the efficacy of feature-based methods. The study also reveals promising results for the 2-Stream Approach with 3D CNN and LSTM, achieving an accuracy of 74%. Furthermore, pretrained models like VGG16 and DenseNet201 along with LSTM exhibit competitive performances with accuracies of 80% each.These results offer valuable insights into the performance of diverse algorithms on the “Coer_Egovision” dataset, providing valuable guidance for future research in the field of egocentric action recognition.