Pedestrian Head and Body Pose Estimation with CNN in the Context of Automated Driving

The challenge of determining pedestrians head poses in camera images is a topic that has already been researched extensively. With the ever-increasing level of automation in the field of Advanced Driver Assistance Systems, a robust head orientation detection is becoming more and more important for pedestrian safety. The fact that this topic is still relevant, however, indicates the complexity of this task. Recently, trained classifiers for discretized head poses have recorded the best results. But large databases, which are essential for an appropriate training of neural networks meeting the special requirements of automatic driving, can hardly be found. Therefore, this paper presents a framework with which reference measurements of head and upper body poses for the generation of training data can be carried out. This data is used to train a convolutional neural network for classifying head and upper body poses. The result is extended in a semi-supervised manner which optimizes and generalizes the detector, so that it is applicable to the prediction of pedestrian intention.