Online action detection
Document typeMaster thesis
Rights accessOpen Access
In online detection, the objective is to detect the start of an action in a video stream as soon as it happens. It is an important yet challenging problem. In many realistic scenarios, we need to detect the action before the action is completed. For example, in the autonomous driving system, it is crucial to detect whether the pedestrian is crossing the street well in time in order to make a decision to stop or to reduce the velocity. Online action detection is a very challenging task in many aspects. It is very hard to predict the start of action for three reasons: First, the background is very diverse. Moreover, there is only a few action instance in a very long video. Last but not least, the model only observes part of the action to predict. To address those challenges, we propose a framework for online action detection and simulate experiments on a large-scale untrimmed video dataset. With the proposed method we have obtained very competitive performance. We also proposed a new evaluation metric for online detection models: Point mean Average Precision (Point mAP), a more appropriate metric than the existing evaluation metrics that have been designed for action detection in an offline setting. We have conducted experiments on THUMOS'14 dataset of video analysis where our proposed model achieved the state-of-the-art performance on the online action detection task.
The details of the work will be defined when the student reaches his destination.