The video salient object detection (SOD) is the first step for the devices in the Internet of Things (IoT) to understand the environment around them. The video SOD needs the objects’ motion information in contiguous video frames as well as spatial contrast information from a single video frame. A large number of IoT devices’ computing power is not sufficient to support the existing SOD methods’ expensive computational complexity in emotion estimation, because they might have low hardware configurations (e.g., surveillance camera, and smartphone). In order to model the objects’ motion information efficiently for SOD, we propose an end-to-end video SOD algorithm with an efficient representation of the objects’ motion information. This algorithm contains two major parts: a 3D convolution-based X-shape structure that directly represents the motion information in successive video frames efficiently, and 2D densely connected convolutional neural networks (DenseNet) with pyramid structure to extract the rich spatial contrast information in a single video frame. Our method not only can maintain a small number of parameters as the 2D convolutional neural network but also represents spatiotemporal information uniformly that enables it can be trained end-to-end. We evaluate our proposed method on four benchmark datasets. The results show that our method achieves state-of-the-art performance compared with the other five methods.
展开▼
