Temporal action recognition in untrimmed videos using deep neural networks

In today's world, the vast amount of video data has increased the importance of deep learning in areas such as semantic information extraction and action recognition. Due to the complex and dynamic nature of videos, there is a need for advanced modeling techniques and algorithms. This study investigates the problem of action recognition in videos with the aim of extracting semantic information from the increasing video contents with digital technologies. Many of the existing studies focus on the classification of short videos. Within the scope of the thesis, an original model based on three-dimensional convolutional neural networks and attention mechanism is proposed for the classification of not only short videos but also long videos. This integration enhances the learning process in both short and long videos, enabling accurate action detection. The proposed model focuses on classifying long videos by first identifying potential event boundaries within these videos using a neural network known as region proposal network, and subsequently performing classification on the proposed video segments. Experimental studies carried out on datasets like HMDB, UCF, and ActivityNet have shown that attention mechanisms significantly improve model performance. The proposed model, integrating 3D convolutional neural networks and attention mechanisms, has improved feature extraction and activity detection capabilities from videos. The model's ability to recognize various activity types and video structures was evaluated using the HMDB and UCF datasets for short video clips and the ActivityNet dataset for longer videos. Specifically, in the UCF and HMDB datasets, the model using the Self Attention mechanism achieved high accuracy rates, while in ActivityNet, the Multi-Head Attention mechanism displayed better ability to recognize complex interactions in longer videos. These findings highlight the crucial role of attention mechanisms in extracting semantic information from videos and reveal the potential of deep learning methods in this area. The obtained results clearly indicate the proposed deep learning model's adaptability to different video structures and its capacity for effective information extraction.