Object tracking is crucial for planning safe maneuvers of mobile robots in dynamic environments, in particular for autonomous driving with surrounding traffic participants. Multi-stage processing of sensor measurement data is thereby required to obtain abstracted high-level objects, such as vehicles. This also includes sensor fusion, data association, and temporal filtering. Often, an early-stage object abstraction is performed, which, however, is critical, as it results in information loss regarding the subsequent processing steps. We present a new grid-based object tracking approach that, in contrast, is based on already fused measurement data. The input is thereby pre-processed, without abstracting objects, by the spatial grid cell discretization of a dynamic occupancy grid, which enables a generic multi-sensor detection of moving objects. On the basis of already associated occupied cells, presented in our previous work, this paper investigates the subsequent object state estimation. The object pose and shape estimation thereby benefit from the freespace information contained in the input grid, which is evaluated to determine the current visibility of extracted object parts. An integrated object classification concept further enhances the assumed object size. For a precise dynamic motion state estimation, radar Doppler velocity measurements are integrated into the input data and processed directly on the object-level. Our approach is evaluated with real sensor data in the context of autonomous driving in challenging urban scenarios.

Occupancy grid mapping approaches, especially those that additionally estimate the dynamics, enable a robust and consistent modeling of the local environment in a cell-level representation. But a scene understanding of surrounding traffic participants requires a generalized object-level representation. This work presents an object tracking approach based on dynamic occupancy grids. The association of occupied grid cells with existing object tracks is solved individually on the cell-level without clustering or forming object hypotheses. New object tracks are extracted using a clustering strategy and a velocity variance analysis of neighboring occupied cells to reduce false positives. In order to improve the estimates of the position and size, an object boundary extraction is presented that takes the surrounding free space of the selected box representation into account. Experimental results with real sensor data show the effectiveness of the proposed object tracking approach in challenging urban scenarios with dense traffic.

Two-dimensional occupancy grid mapping is a common approach for environment mapping and sensor data fusion but non-planar environments are still a remaining issue. The ground shape has to be considered in such environments, especially when low obstacles on the road need to be recognized. Elevation maps are a suitable model because the height is not discretized. This paper presents an improved lidar-based approach for elevation mapping that uses particle filters to estimate the height indirectly by a fusion of lower and upper height boundaries. These boundaries are extracted from lidar reflections and ray geometry of different time steps. Furthermore, a tailored interpolation algorithm is presented that takes the statistics of the particle population of each cell into account. A conclusive qualitative and quantitative evaluation highlights the performance of the presented approach.

Tracking and mapping the local environment form the basis of an autonomous vehicle system. They are often realized separately using occupancy grids, which do not require object or shape assumptions, and model-based object tracking algorithms. Many approaches require a binary classification of the sensor measurements into coming from a static or from a dynamic object, as otherwise inconsistencies between the different representations are likely to occur. This paper presents grid-based tracking and mapping (GTAM), a low-level grid-based approach that simultaneously estimates the static and the dynamic environment, their uncertainties, velocities, as well as information about free space. GTAM works on the level of grid cells, rather than creating object hypotheses. A particle filter is used to obtain continuous cell velocity distributions for all obstacles. Continuous evidences in a Dempster-Shafer model are derived without requiring a binary pre-classification of the sensor measurements. Results and evaluations using a vehicle moving in real dynamic street environments demonstrate the performance of the presented approach.

Mapping and tracking in dynamic environments for autonomously-moving robots is still challenging, despite being essential tasks. They are often done separately using occupancy grids and established object tracking algorithms. In this work, an approach is presented that estimates a uniform, low-level, grid-based world model including dynamic and static objects, their uncertainties, as well as their velocities. It does not require existing object tracks to filter out data points not used for creating and updating the map. Nor does it require that measurements can be classified into belonging to a static or to a moving object. Promising results from experiments with an autonomous vehicle equipped with a laser scanner demonstrate the usefulness of the approach.

This work presents a trajectory clustering approach that groups trajectories without the need of manually-tuned distance thresholds. Contrary to trajectory clustering approaches that use continuous, often geometrically-motivated similarity measures, path similarity is binary. Similar to homotopy classes, path equivalence is based on the obstacles in the environment. The goal states are, however, not fixed, but the paths have certain length restrictions. The equivalence is efficiently checked by closing the paths with sampled intermediate trajectories and using point-in-polygon tests. The proposed algorithm has linear complexity in the number of paths for non-overlapping clusters and, under certain assumptions, also in the case of overlapping clusters. Experimental results from an integration into a path-planning-based road course estimation system are shown and compared to a traditional distance-similarity cluster analysis to demonstrate the performance.

Collision checking is the major computational bottleneck for many robot path and motion planning applications, such as for autonomous vehicles, particularly with grid-based environment representations. Apart from collisions, many applications benefit from incorporating costs into planning; cost functions or cost maps are a common tool. Similar to checking a single configuration for collision, evaluating its cost using a grid-based cost map also requires examining every cell under the robot footprint. This work gives theoretical and practical insights on how to efficiently check a large number of configurations for collision and cost. As part of this work, configuration space costs are formulated, which can be seen as generalization of configuration space obstacles allowing a complete configuration check incorporating the robot geometry to be done using a single lookup. Furthermore, this paper presents two efficient algorithms for their calculation: FAMOD, an approximate method based on convolution, which is independent of the size and the shape of the robot mask, and vHGW-360, an exact method based on the van Herk-Gil-Werman morphological dilation algorithm, which can be used if the robot shape is rectangular. Both algorithms were implemented and evaluated on graphics hardware to demonstrate the applicability and benefit to real-time path and motion planning systems.

In this work, a framework for motion prediction of vehicles and safety assessment of traffic scenes is presented. The developed framework can be used for driver assistant systems as well as for autonomous driving applications. In order to assess the safety of the future trajectories of the vehicle, these systems require a prediction of the future motion of all traffic participants. As the traffic participants have a mutual influence on each other, the interaction of them is explicitly considered in this framework, which is inspired by an optimization problem. Taking the mutual influence of traffic participants into account, this framework differs from the existing approaches which consider the interaction only insufficiently, suffering reliability in real traffic scenes. For motion prediction, the collision probability of a vehicle performing a certain maneuver, is computed. Based on the safety evaluation and the assumption that drivers avoid collisions, the prediction is realized. Simulation scenarios and real-world results show the functionality.

The road course is an essential feature for many driver assistance systems and for autonomously-maneuvering vehicles. It is commonly stored in a map and hence assumed to be known a-priori. There are however situations in which the map data can become invalid, such as in road construction sites. In other situations, localization in the map might not be accurate enough, which can happen, for example, in dense urban areas. In this work, a novel approach to road course estimation is presented that is based on path planning through grid maps under non-holonomic and velocity constraints. With this approach, it is possible to estimate the road boundaries on a wide range of roads, including roads with continuous as well as discontinuous borders, roads exhibiting strong curvatures or S-shapes and road junctions. Furthermore, a plausibility measure is given to validate the road course and it is shown how the road center can be smoothed.