Pose Recognition for Dense Vehicles under Complex Street Scenario

Abstract

Locating vehicles in the surrounding environment and distinguishing their poses are crucial in autonomous driving systems. An intelligent transportation system expects a vision-based object detector not only to detect and classify vehicles but also to recognize the poses of them in estimating their future movement direction. However, most of the current object detection algorithms only classify different classes of vehicles; and there are rare studies on classifying different poses within the same class of vehicles. In addition, the currently available datasets have no annotation data that are specific to different poses of intra-class vehicles. To achieve the vehicles pose recognition, we build a vehicle pose dataset based on the Cityscapes dataset. Furthermore, we also improve the network model of the YOLO v2 detector to support simultaneous vehicle detection and pose classification with an improvement of 10.71% mean average precision (mAP) comparing to the default network which runs in real-time with 29 FPS.