Real-Time Motion Detection Network Based on Single Linear Bottleneck and Pooling Compensation

Motion (change) detection is a basic preprocessing step in video processing, which has many application scenarios. One challenge that deep learning-based methods require high computation power to improve their accuracy. In this paper, we introduce novel semantic segmentation and lightweight-based network for motion detection, called Real-time Detection Network Based on Single Linear Bottleneck Pooling Compensation (MDNet-LBPC). the feature extraction stage, most computationally expensive CNN block replaced with our single linear bottleneck operator reduce computational cost. During decoder pooling compensation mechanism can supplement useful information. To best knowledge, first work use lightweight solve task. We show acceleration performance of 5% higher than bottleneck, more suitable improving efficiency model inference. On dataset CDNet2014, MDNet-LBPC increases frames per second (FPS) metric by 123 compared suboptimal method FgSegNet_v2, ranking inference speed. Meanwhile, achieves 95.74% accuracy metric, comparable state-of-the-art methods.