Automotive Interior Sensing - Temporal Consistent Human Body Pose Estimation

Abstract

With the emergence and development of autonomous vehicles, a necessity to constantly monitor and identify objects and action that occur in the surrounding environment of the vehicle itself was also created. This type of monitoring is particularly important in the case of shared vehicles, given the necessity to identify actions not only in the exterior but also in the interior of the vehicle due to the absence of a human driver that can detect, for instance, potential violent actions between passengers and/or cases where assistence is required. Encompassed in this context, Bosch has developed a human body pose estimation solution in order to extrapolate the pose of all vehicle occupants present in a given image, infere the behaviour of each passenger and, consequently, identify potentially malicious actions. However, in order to apply this algorithm not only to isolated images but also to videos it is necessary to add temporal context between frames. In other words, an association is required between the body pose estimation for a given person in a given frame and the body pose estimations for the same person in subsequent frames in order to ensure that the identification of that passenger (or any other passenger present in the same frame) is accurate and consistent throughout the entire video. The temporal association topic, also known as pose tracking, is addressed and developed during the present project, culminating in the proposal and implementation of a solution that considerably improves the temporal consistency of the human body pose estimation algorithm developed by Bosch. The implemented solution uses a mixture of currently relevant classical approaches for data association, such as the Hungarian algorithm e Intersection over Union techniques, and approaches based on data logic developed specifically for the present case. Regarding performance, the developed algorithm is evaluated using two of the most recurrent metrics for pose tracking methods.