Smoke Detection on Video Sequences Using Convolutional and Recurrent Neural Networks

The combination of a convolutional neural network (CNN) and recurrent neural network (RNN) is proposed to detect the smoke in space and time domains. CNN part automatically builds the low-level features, and RNN part finds the relation between the features in different frames of the same event. For this work, the new dataset was constructed with at least 64 sequential frames for each set giving the network ability to analyze the behavior of the smoke for at least 2 s. While being not too deep thus allowing fast processing, the proposed network outperformed state of the art deep CNNs which do not consider the change of the object in time.