Velocities measured by radar sensors suffer from ambiguities caused by aliasing effects associated with signal processing. These ambiguities are highly undesired as they affect otherwise very accurate speed measurements. For automotive applications utilizing radar sensors, this means that measured velocities are potentially unreliable which can lead to unsafe conditions for automated driving functionalities. This work presents the first approach to disambiguate radial velocities measured by radar sensors based on Deep Learning methods. By utilizing the presented method, radial velocity estimates can be obtained which are both highly accurate while being reliable as ambiguities are resolved.

We propose a combination of optical tracker, Probability Hypothesis Density (PHD) Filter and group tracking for tracking vehicle head lights and tail lights from an in vehicle, forward facing camera. We propose these systems are advantageous, because they can bridge several frames without outside detections and lead to more stable tracks than just using a traditional tracker, like a Kalman filter, on it's own. Additionally PHD does not need track-data association, but moves association uncertainty into the tracker, where it can be incorporated in covariance and noise calculations. Evaluation is performed with a closed source detector and a private data set. This evaluation proves the stability of the tracks and the trackers ability to bridge large amounts of time without external detections. This makes it a suitable choice for high difficulty situations that lead to the external detector missing light sources.

In current state-of-the-art systems for object detection and classification a huge amount of data is needed. Even if large databases are available, some classes are typically underrepresented and therefore the classifier is not able to capture the variability in appearance. In this work we present a novel method to enrich the training database with natural looking synthetic images. The method can be used to transfer the object appearance from one image (template image) to another image (base image) containing a different object of the same or a similar category. In order to preserve natural appearance and avoid artifacts we only use the gray-level values of the base image for synthesis. The main contribution of this work is an extension of the shift-map approach [1]. An appropriate optimization criteria for the used Markov Random Field (MRF) is defined and the MRF is embedded into a general framework for training data synthesis, which is exemplary tailored to the generation of traffic signs. The influence of using synthetic images is evaluated using a convolutional neural network (CNN).

Measuring absolute distances in monocular vision is challenging and cannot be solved directly. Conventionally, assumptions like an a priori width of the target object or geometric constraints are made to overcome the problem.

Knowledge about the host egomotion can help to stabilize and improve many applications in the advanced driver assistance domain. It can be a crucial feature for object tracking and calibration. In this paper we describe a novel approach which is fast to compute and robust. We utilize a depth prior for the translation and integrate robust estimation techniques, like MSAC and an M-estimator. The MSAC is further improved by imposing prior information directly into the MSAC step. We can show that using this scheme is fast and enhances our results. For testing we utilize a large video dataset from which we also have computed the pose estimates via sparse bundle adjustment. Using a loop-closing sequence we also qualitatively analyze our results. The presented approach has been tested online on a car PC and as such can be computed in real time.

This paper focuses on dynamic orientation estimation of a vehicle-mounted mono camera. In particular, the pitch and roll angles of the camera relative to the road surface. Information about the orientation angles is included in the homography (projective transformation) between two images of a planar road surface. The extraction of angles from a homography matrix is possible but not recommended due to parameter ambiguities. For this reason we do not estimate a full homography matrix but reduce the parameter space to two parameters. In this area of parameter estimation there are mainly two different approaches: The optical flow based approach and the image registration based approach. In order to decide which of these approaches is more favorable for the angle estimation problem, we develop one optical flow based and one image registration based angle estimation algorithm. We are the first directly evaluating and comparing both approaches with each other with the help of an artificial image sequence as well as real-world driving scenarios. In addition, this paper lifts the common limitation of roll angle of zero degree for dynamic camera orientation estimation. Our research finds that there is only a small difference between the parameter estimation results of the optical flow and image registration based approach.

In this paper a novel spline-based multi-lane detection and tracking system is proposed. Reliable lane detection and tracking is an important component of lane departure warning systems, lane keeping support systems or lane change assistance systems. The major novelty of the proposed approach is the usage of the so-called Catmull-Rom spline in combination with the extended Kalman filter tracking. The new spline-based model enables an accurate and flexible modeling of the lane markings. At the same time the application of the extended Kalman filter contributes significantly to the system robustness and stability. There is no assumption about the parallelism or the shapes of the lane markings in our method. The number of lane markings is also not restrained, instead each lane marking is separately modeled and tracked. The system runs on a standard PC in real time (i.e. 30 fps) with WVGA image resolution (752 × 480). The test vehicle has been driven on the roads with challenging scenarios, like worn out lane markings, construction sites, narrow corners, exits and entries of the highways, etc., and good performance has been demonstrated. The quantitative evaluation has been performed using manually annotated video sequences.

This paper presents an automotive video and radar data fusion framework that can be used as a preliminary stage of an automatic cruise control or collision mitigation by braking system. The fusion framework finds the optimal assignment of radar and camera target reports and provides improved state estimates for the fused targets. A sophisticated critical path selection is presented and used in the critical target selection module that aims to select the most relevant target. This module is capable of identifying targets that cut into the ego lane or cut out from the ego lane and incorporate that into the final target selection. The selected target is then compared to a state of the art algorithm within the radar sensor. Additional test drives were made to evaluate the performance of the new algorithm. Due to its low computational effort and the sensor independent design the presented algorithm is suitable to be used in the automotive embedded environment.

A novel approach for a decision fusion and reasoning system for vision-based traffic sign recognition is presented. This module consists of several steps. In the first stage, a track-based Bayesian fusion scheme is used to fuse the classification results from each frame to obtain a fusion result for each track to decide whether a sign is present, as well as to determine the sign type. In order to determine the sign type, the temporal fusion scheme has been combined with a decision tree. In the second stage, the system combines and fuses probable identical objects which help to further reduce failures in the recognition process. The decision is based on the fusion results, as well as a position cue. Finally, a reasoning module is used to decide which of the passed signs should be shown to the driver. In addition to these modules, a general evaluation method for multi-class tracking systems is shown. While some failures are observed from the evaluation on object level, the additional post processing steps improve the system in such a way that the finally presented signs are almost always correct on the test set.