서강대학교

초록/요약

For researches on water detection in CCTV videos, it is essential to have methods that extract various features of water and datasets that contain features of water. Existing methods for water detection do not take into account the various features of water, making it difficult to operate properly in CCTV videos. In this thesis, we propose a network for detecting water in CCTV videos and a video dataset to include various spatio-temporal features of water. The proposed network uses U-net with the encoder transformed into a two-stream structure. Each stream encoder extracts spatial and temporal features of water using 2D and temporal 3D convolution. The temporal 3D convolution can achieve reasonable performance with feasible computation than 3D convolution. The decoder of the proposed network receives the concatenated feature map of each stream through skip connections. Finally, the water is detected through the softmax function. In network training, we use the combined Dice loss and L1 loss to give a larger gradient for false prediction. Throughout all experiments, we use the training and testing sets of proposed dataset. Experimental results show that the temporal features of water improve the water detection performance. The proposed network shows reasonable performance with less computation than the network using 3D convolution. Also, the proposed network with pre-trained model performs better than the existing semantic segmentation networks. In particular, the water is well detected even in areas that are difficult to distinguish from surrounding objects because of water reflection and object shadow.