Convolutional Neural Network for Finding Unusual Frames in Outdoor Webcam Streams
Kanghui Liu · Western CEDAR (Western Washington University) · 2020
Outdoor live streaming webcams produce vast quantities of video footage that reveal rich visual information about scenes around the world and how they change. The ability to identify unique or unusual frames in these streams has potential applications in surveillance, photography, and traffic monitoring, but the volume of data makes manual analysis impractical. Existing techniques tend to focus on detecting unusual or suspicious events in a surveillance context, requiring high-level object motion analysis and rely on hand-crafted features. In this paper, we propose a deep learning-based model that uses Convolutional Neural Networks(CNNs) to solve the novel proxy task of predicting the time interval between frames, using the model’s inaccurate predictions as an indicator that something unexpected has happened in the scene. We evaluate our techniques on a new dataset of live-streaming outdoor webcam footage. Experimental results on this dataset confirm that our proposed model is capable of finding unusual frames, with 1% ~ 10% of the frames from the input videos are labeled as unusual.