Next-generation automotive radars provide elevation data in addition to range-, azimuth- and Doppler velocity. In this experimental study, we apply a state-of-the-art object detector (PointPillars), previously used for LiDAR 3D data, such 3+1D radar (where 1D refers Doppler). ablation studies, first explore the benefits of additional information, together with that Doppler, cross section temporal accumulation, context multi-class road user detection. We subsequently compare detection...

Abstract Visual place recognition (VPR) is the process of recognising a previously visited using visual information, often under varying appearance conditions and viewpoint changes with computational constraints. VPR related to concepts localisation, loop closure, image retrieval critical component many autonomous navigation systems ranging from vehicles drones computer vision systems. While concept has been around for years, research grown rapidly as field over past decade due improving...

Anticipating future situations from streaming sensor data is a key perception challenge for mobile robotics and automated vehicles. We address the problem of predicting path objects with multiple dynamic modes. The dynamics such targets can be described by Switching Linear Dynamical System (SLDS). However, predictions this probabilistic model cannot anticipate when change in mode will occur. propose to extract various types cues computer vision provide context on target's behavior,...

This letter presents a novel radar based, single-frame, multi-class detection method for moving road users (pedestrian, cyclist, car), which utilizes low-level cube data. The provides class information both on the target- and object-level. Radar targets are classified individually after extending target features with cropped block of 3D around their positions, thereby capturing motion parts in local velocity distribution. A Convolutional Neural Network (CNN) is proposed this classification...

We present a novel open-source tool for extrinsic calibration of radar, camera and lidar. Unlike currently available offerings, our facilitates joint all three sensing modalities on multiple measurements. Furthermore, target design extends existing work to obtain simultaneous measurements these modalities. study how various factors the procedure affect outcome real multi-modal target. Three different configurations optimization criterion are considered, namely using error terms minimal...

We address joint extrinsic calibration of lidar, camera and radar sensors. To simplify calibration, we propose a single target design for all three modalities, implement our approach in an open-source tool with bindings to Robot Operating System (ROS). Our features optimization configurations, namely using error terms minimal number sensor pairs, or pairs combination loop closure constraints, by adding structure estimation probabilistic model. Apart from relative where transformations...

We present a probabilistic framework for the joint estimation of pedestrian head and body orientation from mobile stereo vision platform. For both parts, we convert responses set orientation-specific detectors into (continuous) probability density function. The parts are localized by means <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><i>pictorial structure</i></b> approach, which balances part-based detector with spatial constraints. Head...

We learn motion models for cyclist path prediction on real-world tracks obtained from a moving vehicle, and propose to exploit the local road topology obtain better predictive distributions. The are extracted Tsinghua-Daimler Cyclist Benchmark detection, corrected vehicle egomotion. Tracks then spatially aligned curves crossings in road. study standard approach literature based Kalman Filters, as well mixture of specialized filters related specific orientations at junctions. Our experiments...

Abstract Conventional decision trees have a number of favorable properties, including small computational footprint, interpretability, and the ability to learn from little training data. However, they lack key quality that has helped fuel deep learning revolution: being end-to-end trainable. Kontschieder et al. (ICCV, 2015) addressed this deficit, but at cost losing main attractive trait trees: fact each sample is routed along subset tree nodes only. We here present an scheme for...

This work addresses cross-view camera pose estimation, i.e., determining the 3-Degrees-of-Freedom of a given ground-level image w.r.t. an aerial local area. We propose SliceMatch, which consists ground and feature extractors, aggregators, predictor. The extractors extract dense features from images. Given set candidate poses, aggregators construct single descriptor pose-dependent descriptors. Notably, our novel aggregator has attention module for ground-view guided selection utilizes...

This paper proposes a Recurrent Neural Network (RNN) for cyclist path prediction to learn the effect of contextual cues on behavior directly in an end- to-end approach, removing need any annotations. The proposed RNN incorporates three distinct cues: one related actions cyclist, location road, and interaction between egovehicle. predicts Gaussian distribution over future position second into with higher accuracy, compared current state-of-the-art model that is based dynamic mode annotations,...

This paper presents our research platform Safe VRU for the interaction of self-driving vehicles with Vulnerable Road Users (VRUs, i.e., pedestrians and cyclists). The details design (implemented a modular structure within ROS) full stack vehicle localization, environment perception, motion planning, control, emphasis on perception planning modules. detects VRUs using stereo camera predicts their paths Dynamic Bayesian Networks (DBNs), which can account switching dynamics. planner is based...

Early and accurate detection of crossing pedestrians is crucial in automated driving order to perform timely emergency manoeuvres. However, this a difficult task urban scenarios where are often occluded (not visible) behind objects, e.g., other parked vehicles. We propose an occlusion aware fusion stereo camera radar sensors address with such Our proposed method adapts both the expected rate properties detections different areas according visibility sensors. In our experiments on real-world...

Accurate motion models are key to many tasks in the intelligent vehicle domain, but simple Linear Dynamics (e.g. Kalman filtering) do not exploit spatio-temporal context of motion. We present a method learn Switching object tracks observed from within driving vehicle. Each switching state captures dynamics as mean with variance, also has an additional spatial distribution on where dynamic is seen relative Thus, both object's previous movements and current location will make certain more...

We present an approach for the joint probabilistic estimation of pedestrian head and body orientation in context intelligent vehicles. For both, body, we convert output a set orientation-specific detectors into full (continuous) probability density function. The parts are localized with pictorial structure which balances part-based detector spatial constraints. Head estimates furthermore coupled probabilistically to account anatomical Finally, single-frame integrated over time by particle...

Early and accurate detection of crossing pedestrians is crucial in automated driving to execute emergency manoeuvres time. This a challenging task urban scenarios however, where people are often occluded (not visible) behind objects, e.g. other parked vehicles. In this paper, an occlusion aware multi-modal sensor fusion system proposed address with Our method adjusts the rate different areas based on visibility. We argue that using information can help evaluate measurements. experiments real...

This work proposes to use passive acoustic perception as an additional sensing modality for intelligent vehicles. We demonstrate that approaching vehicles behind blind corners can be detected by sound before such enter in line-of-sight. have equipped a research vehicle with roof-mounted microphone array, and show on data collected this sensor setup wall reflections provide information the presence direction of occluded A novel method is presented classify if from what it visible, using input...

We propose a novel end-to-end method for cross-view pose estimation. Given ground-level query image and an aerial that covers the query's local neighborhood, 3 Degrees-of-Freedom camera of is estimated by matching its descriptor to descriptors regions within image. The orientation-aware are obtained using translationally equivariant convolutional ground encoder contrastive learning. Localization Decoder produces dense probability distribution in coarse-to-fine manner with Matching Upsampling...