MOSnet: Moving object segmentation with convolutional networks

Identifying moving objects is considered a difficult problem owing to camera motion, motion blur and appearance changes. To solve these problems, a moving object segmentation method based on a convolutional neural network is presented. The proposed network takes successive image pairs as input, and predicts the per-pixel motion status. This process consists of three streams: One that learns appearance features, another that learns motion features and a third that combines both features. Therefore, a joint model is learned for segmenting a moving object, because appearance and motion features complement each other. Experimental results, based on a challenging dataset, demonstrate that the proposed method has superior performance over stateof- the-art methods, with respect to intersection over union and F-measure scores.