3D convolutional neural networks for Online Detection of Action Start

Online action detection refers to detect an action as it happens and ideally even before the action is fully completed. In this thesis, we focus on online action detection in untrimmed videos. This is important because videos in many applications such as surveillance monitoring system needs to detect anomaly actions as soon as possible and then issue an alert to allow timely security response. I developed a convolutional network to address this problem. The network is coded in Keras (Tensorflow backend) framework and based on C3D network. It is trained with three methods: (1) use adaptive sampling to handle the scarcity action start samples problem; (2) model the temporal consistency to make the feature of action start window close to the actual action; (3) finetuning the model via Generative adversarial network. I conduct experiments using THUMOS’14 dataset and use a point-level mAP to evaluate the results.