A Fast Multi-Task CNN for Spatial Understanding of Traffic Scenes

Traffic scene understanding involves the perception and accurate representation of a vehicles surroundings within a comprehensive environment model. Generally, automotive environment models incorporate multiple different and complementary representations each focusing on specific aspects of traffic scenes such as the drivable terrain, dynamic objects, the lane layout, etc. Considering camera based environment perception, multi-task models offer an efficient approach capable of generating complementary representations in an integrated manner. Towards this objective, we propose a convolutional neural network (CNN) comprised of a shared encoder stage and specific decoders for road segmentation and object detection. In comparison to exclusive models for each separate task, the redundant computation of common feature maps is avoided resulting in fast inference times even on low cost embedded systems. Furthermore, the detection stage is extended to predict the orientation of detected objects. The orientation estimate guides a consecutive 3D bounding box estimation based on analytic geometry. Thus, the CNN suffices to generate a basic environment model consisting of the spatial information of dynamic objects as well as the free space area obtained from road segmentation.