To verify and validate networks, it is essential to gain insight into their decisions, limitations as well as possible shortcomings of training data. In this work, we propose a post-hoc, optimization ...based visual explanation method, which highlights the evidence in the input image for a specific prediction. Our approach is based on a novel technique to defend against adversarial evidence (i.e. faulty evidence due to artefacts) by filtering gradients during optimization. The defense does not depend on human-tuned parameters. It enables explanations which are both fine-grained and preserve the characteristics of images, such as edges and colors. The explanations are interpretable, suited for visualizing detailed evidence and can be tested as they are valid model inputs. We qualitatively and quantitatively evaluate our approach on a multitude of models and datasets.
This paper addresses the problem of path planning for a non-holonomic vehicle in fast changing urban environments. The developed method is flexible in utilization and generates collision-free paths ...with continuous curvature for a wide range of maneuvers such as regular lane-driving, rectangular or parallel parking and K-Turns. An A* algorithm is used as incremental planning method in combination with a kinematic model to follow non-holonomic constraints. A shortest-path estimate is used as optimization criteria. In the context of collision-free driving, the idea of using risk maps for representation of obstacles and other environmental influences is presented. Simulation and experiments show the applicability of this approach for lane driving as well as narrow navigation in parking spots. Further ideas of enhancements are discussed in the outlook of the paper.
This paper presents a generic approach for tactical decision-making under uncertainty in the context of driving. The complexity of this task mainly stems from the fact that rational decision-making ...in this context must consider several sources of uncertainty: The temporal evolution of situations cannot be predicted without uncertainty because other road users behave stochastically and their goals and plans cannot be measured. Even more important, road users are only able to perceive a tiny part of the current situation with their sensors because measurements are noisy and most of the environment is occluded. In order to anticipate the consequences of decisions a probabilistic approach, considering both forms of uncertainty, is necessary. We address this by formulating the task of driving as a continuous Partially Observable Markov Decision Process (POMDP) that can be automatically optimized for different scenarios. As driving is a continuous-space problem, the belief space is infinite-dimensional. We do not use a symbolic representation or discretize the state space a priori because there is no representation of the state space that is optimal for every situation. Instead, we employ a continuous POMDP solver that learns a good representation of the specific situation.
Building a model of the environment is essential for mobile robotics. It allows the robot to reason about its surroundings and plan actions according to its intentions. To enable safe motion planning ...it is vital to anticipate object movements. This paper presents an improved formulation for occupancy filtering. Our approach is closely related to the Bayesian Occupancy Filter (BOF) presented in. The basic idea of occupancy filters is to represent the environment as a 2-dimensional grid of cells holding information about their state of occupancy and velocity. To improve the accuracy of predictions, prior knowledge about the motion preferences is used, derived from map data that can be obtained from navigation systems. In combination with a physically accurate transition model, it is possible to estimate the environment dynamics. Experiments show that this yields reliable estimates even for occluded regions.
Bayesian Occupancy Filtering is an alternative to classical object tracking. Instead of estimating the state of objects in the environment, the latter is separated into equidistant cells. Tracking ...the occupancy state of these grid-cells is sufficient for many applications in robotics and cell-measurements can be easily produced from almost any kind of sensor. In 6 a sophisticated occupancy filter named BOFUM (Bayesian Occupancy Tracking using prior Map Knowledge) is introduced, which is able to infer velocities solely from occupancy measurements. It also features an advanced process model with motion uncertainty, which can be specialized for different application needs. In this paper we present an approach for recursively applying importance sampling (IS) to approximate the BOFUM calculations. The approach is similar to well known particle filters, but for a discrete cell perspective. In our experiments we achieved a speedup of at least 40-times by using the IS, thus making the algorithm applicable in real-world applications. We evaluate the consequences of approximation in an urban traffic scenario and also show the drawbacks of sampling.
This paper presents a filter that is able to simultaneously estimate the behaviors of traffic participants and anticipate their future trajectories. This is achieved by recognizing the type of ...situation derived from the local situational context, which subsumes all information relevant for the drivers decision making. By explicitly taking into account the interactions between vehicles, it achieves a comprehensive situational understanding, inevitable for autonomous vehicles and driver assistance systems. This provides the necessary information for safe behavior decision making or motion planning. The filter is modeled as a Dynamic Bayesian Network. The factored state space, modeling the causal dependencies, allows to describe the models in a compact fashion and reduces the computational complexity of the inference process. The filter is evaluated in the context of a highway scenario, showing a good performance even with very noisy measurements. The presented framework is intended to be used in traffic environments but can be easily transferred to other robotic domains.
This paper reports on the behaviour decision and execution unit of AnnieWAY , an autonomous vehicle that is capable of driving through urban scenarios and that has successfully entered the finals of ...the DARPA Urban Challenge 2007 competition. Starting with a short description of the car and its major hardware components, we outline the underlying software structure and focus on the design of the behavior decision module. The selection of maneuvers necessary to accomplish the mission is conducted online via a concurrent hierarchical state machine that specifically ascertains behavior in accordance with California traffic rules. The states and transitions used to model a adequate behaviour are described. We conclude with a report of the results achieved during the competition.
This paper describes an algorithm for handling moving traffic which was deployed on AnnieWAY, an autonomous vehicle successfully entering the finals of the DARPA Urban Challenge 2007 competition. The ...algorithm allows for a robust and effective collision check for a variety of maneuvers including turning at intersections with oncoming traffic, merging into moving traffic, changing lanes, as well as dynamic passing and can easily be integrated into a high-level decision process, such as a state machine.
Probabilistic MDP-behavior planning for cars Brechtel, Sebastian; Gindele, Tobias; Dillmann, Rudiger
2011 14th International IEEE Conference on Intelligent Transportation Systems (ITSC),
2011-Oct.
Conference Proceeding
This paper presents a method for high-level decision making in traffic environments. In contrast to the usual approach of modeling decision policies by hand, a Markov Decision Process (MDP) is ...employed to plan the optimal policy by assessing the outcomes of actions. Using probability theory, decisions are deduced automatically from the knowledge about how road users behave over time. This approach does neither depend on an explicit situation recognition nor is it limited to only a variety of situations or types of descriptions. Hence it is versatile and powerful. The contribution of this paper is a mathematical framework to derive abstract symbolic states from complex continuous temporal models encoded as Dynamic Bayesian Networks (DBN). For this purpose discrete MDP states are interpreted by random variables. To make computation feasible this space grows adaptively during planning and according to the problem to be solved.