Latent Markov Chain Encoding for Abnormal Landing Event Detection

Landing is a high-risk stage compared to other flight stages. The goal of this research is to detect abnormal landing events at early phase from live video surveillance alone. A weakly supervised feature coding method is proposed to enforce features shift in a predictable pattern. On top of this, a latent Markov chain model is proposed for abnormal events detection. Finally, a novel recurrent neural network (RNN) with relatively simple structure is proposed which integrates all proposed methods into one end-to-end model. Experiment results shows that this model is able to generate interpretable features from intermediate layers. The proposed model is capable to provided competitive abnormal event detection results.