Learning to Analyze Visual Data Streams for Environment Perception

Sammanfattning: A mobile robot, instructed by a human operator, acts in an environment with many other objects. However, for an autonomous robot, human instructions should be minimal and only high-level instructions, such as the ultimate task or destination. In order to increase the level of autonomy, it has become a foremost objective to mimic human vision using neural networks that take a stream of images as input and learn a specific computer vision task from large amounts of data. In this thesis, we explore several different models for surround sensing, each of which contributes to a higher understanding of the environment being possible. As its first contribution, this thesis presents an object tracking method for video sequences, which is a crucial component in a perception system. This method predicts a fine-grained mask to separate the pixels corresponding to the target from those corresponding to the background. Rather than tracking location and size, the method tracks the initial pixels assigned to the target in this so-called video object segmentation. For subsequent time steps, the goal is to learn how the target looks using features from a neural network. We named our method A-GAME, based on the generative modeling of deep feature space, separating target and background appearances. In the second contribution of this thesis, we detect, track, and segment all objects from a set of predefined object classes. This information is how the robot increases its capabilities to perceive the surroundings. We experiment with a graph neural network to weigh all new detections and existing tracks. This model outperforms prior works by separating visually, and semantically similar objects frame by frame. The third contribution investigates one limitation of anchor-based detectors, which classify pre-defined bounding boxes as either negative or positive and thus provide a limited set of handled object shapes. One idea is to learn an alternative instance representation. We experiment with a neural network that predicts the distance to the nearest object contour in different directions from each pixel. The network then computes an approximated signed distance function containing the respective instance information. Last, this thesis studies a concept within model validation. We observed that overfitting could increase performance on benchmarks. However, this opportunity is insipid for sensing systems in practice since measurements, such as length or angles, are quantities that explain the environment. The fourth contribution of this thesis is an extended validation technique for camera calibration. This technique uses a statistical model for each error difference between an observed value and a corresponding prediction of the projective model. We compute a test over the differences and detect if the projective model is incorrect.