Multi-Channel Fused Lasso for Motion Detection in Dynamic Video Scenarios

Rong Gao, Xin Liu, Jingyu Yang, Huanjing Yue

Abstract: Motion detection is a fundamental step in analyzing video sequences, capable of enhancing consumer electronics products with increased intelligence, interactivity, and convenience. Structured and fused sparsity has been used in previous works to normalize the foreground signal due to the foreground’s spatial and temporal coherence. As far as we are aware, no previous works have studied the group prior to multi-channels (such as the RGB) to the foreground signals. However, a multi-channel signal is the correct representation of a pixel. Under the condition that one pixel is equal (similar) to its neighboring pixels, it’s reasonable that the three channels of RGB should also be identical (similar). This work investigates the smoothness of multi-channel signals by proposing a novel regularizer named the Multi-Channel Fused Lasso (MCFL). Specifically, we introduce a two-channel structure to implement motion detection. First, low-rank matrix decomposition is performed on the video footage along different planes. Low-rank background and sparse foreground (rough foreground candidate for the second pass) are segmented from the video sequence. Further, MCFL regularization is used for sparse signal recovery to improve the performance of the foreground mask. The proposed method is validated on different challenging videos. Sufficient experimental results show that our method is effective in a variety of challenging scenarios. Compared with the current best sparsely-based method, the performance of F-Measure improves by 0.4, 0.4, and 0.1 respectively on the I2R, BMC, and CDnet2014 datasets. Our approach is also competitive compared to the deep learning models. Our code can be obtained at https://github.com/linuxsino/MCFL .