Spatio-Temporal Neural Network with Dilated Retrospective Convolution for Short-Time Change Detection

• Published in: 2019 7th International Conference on Information, Communication and Networks (ICICN) pp. 213 - 218
• Main Authors: Chen, Chao, Zhang, Sheng
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.04.2019
• Subjects: Convolution; dilated retrospective convolution; Feature extraction; Kernel; Neural networks; short-time change detection; Task analysis; Testing; Training
• ISBN: 1728104254; 9781728104256
• DOI: 10.1109/ICICN.2019.8834948

Change detection remains to be challenging due to variability of real-world conditions. Existing methods usually require a long-time observation, from tens to hundreds of video frames as reference, to build a robust background model for separating changing foreground from background. The performances of these methods decline severely when subject to quite fewer reference frames. In this paper, we focus on the task of short-time change detection, which aims to detect changing foreground with only a few preceding reference frames. We exploit the capability of spatial-temporal neural network in learning and representing change-cued information, and propose a novel dilated retrospective convolution which enables feature extraction between the current frame and all reference frames in a retrospective manner, and uses spatially dilated kernel for flexibly expanded inter-frame field-of-views. Trained end-to-end on hard samples from various context, our architecture performs accurate in perceiving changes and effective in combating noises in complex scenarios. Extensive evaluation on CDnet demonstrates substantial superiority of our proposed method.