Pose Based Action Recognition of Vulnerable Road Users Using Recurrent Neural Networks

This work investigates the use of knowledge about three dimensional (3D) poses and Recurrent Neural Networks (RNNs) for detection of basic movements, such as wait, start, move, stop, turn left, turn right, and no turn, of pedestrians and cyclists in road traffic. The 3D poses model the posture of individual body parts of these vulnerable road users (VRUs). Fields of application for this technology are, for example, driver assistance systems or autonomous driving functions of vehicles. In road traffic, VRUs are often occluded and only become visible in the immediate vicinity of the vehicle. Hence, our proposed approach is able to classify basic movements after different and especially short observation periods. The classification will then be successively improved in case of a longer observation. This allows countermeasures, such as emergency braking, to be initiated early if necessary. The benefits of using 3D poses are evaluated by a comparison with a method based solely on the head trajectory. We also investigate the effects of different observation periods. Overall, knowledge about 3D poses improves the basic movement detection, in particular for short observation periods. The greatest improvements are achieved for the basic movements start, stop, turn left, and turn right.