Driver Anomaly Detection Using Skeleton Images

Abstract

Many unexpected situations can occur while driving that may lead to dangerous accidents. Some of them may be caused by sudden health problems (e.g. heart attack, stroke, total collapse) or by driver inattention (e.g. microsleep, visual distraction). This has motivated the need for developing the methods that are able to monitor the driver’s state in the first step and to prevent the accidents in the second step (e.g. by activating an acoustic signal, or even by taking over driving). In this paper, we propose a method that can be used for detecting the abnormal driving situations. Our approach is based on two main steps. In the first step, the MNIST-like skeleton images are created with the use of human pose detector. In the second step, an appropriate neural network is used for the final classification. Since we also include the anomalies consisting in an unusual trajectory of a certain body part (not only an unusual shape of body, which can be detected from the isolated images), short sequences of images are examined. The LSTM (long short-term memory) autoencoder is used as a main network architecture. The experiments that are presented show that the proposed method achieves better results than other compared methods.