Human Activity Recognition and Abnormality Detection Using Deep Learning
Chaitanya Maradana, Min Kyung An, Amar A. Rasheed, Qingzhong Liu · 2024
Video classification is a crucial task in computer vision with applications ranging from surveillance to sports analytics. This paper presents a comprehensive investigation into the fine-tuning of VideoMAE, a state-of-the-art model originally trained on the Kinetics 400 dataset, for improved performance across various video classification domains. Our study focuses on adapting VideoMAE to four distinct datasets: UCF Crime, UCF Crime (anomaly detection only), Action Recognition, and Sports Recognition. Through meticulous fine-tuning procedures and experimental evaluations, we analyze the efficiency of transfer learning from Kinetics 400 to these specialized domains. We explore the nuances of each dataset, addressing challenges such as class imbalance, anomaly detection, and diverse motion patterns. Furthermore, we showcase the accuracy and loss plots for comparison of the different datasets.