Point Cloud Painting for 3D Object Detection with Camera and Automotive 3+1D RADAR Fusion.

RADARs and cameras have been present in automotives since the advent of ADAS, as they possess complementary strengths and weaknesses but have been underlooked in the context of learning-based methods. In this work, we propose a method to perform object detection in autonomous driving based on a geometrical and sequential sensor fusion of 3+1D RADAR and semantics extracted from camera data through point cloud painting from the perspective view. To achieve this objective, we adapt PointPainting from the LiDAR and camera domains to the sensors mentioned above. We first apply YOLOv8-seg to obtain instance segmentation masks and project their results to the point cloud. As a refinement stage, we design a set of heuristic rules to minimize the propagation of errors from the segmentation to the detection stage. Our pipeline concludes by applying PointPillars as an object detection network to the painted RADAR point cloud. We validate our approach in the novel View of Delft dataset, which includes 3+1D RADAR data sequences in urban environments. Experimental results show that this fusion is also suitable for RADAR and cameras as we obtain a significant improvement over the RADAR-only baseline, increasing mAP from 41.18 to 52.67 (+27.9%).

Reinforcement Learning-Based Autonomous Driving at Intersections in CARLA Simulator.

Intersections are considered one of the most complex scenarios in a self-driving framework due to the uncertainty in the behaviors of surrounding vehicles and the different types of scenarios that can be found. To deal with this problem, we provide a Deep Reinforcement Learning approach for intersection handling, which is combined with Curriculum Learning to improve the training process. The state space is defined by two vectors, containing adversaries and ego vehicle information. We define a features extractor module and an actor-critic approach combined with Curriculum Learning techniques, adding complexity to the environment by increasing the number of vehicles. In order to address a complete autonomous driving system, a hybrid architecture is proposed. The operative level generates the driving commands, the strategy level defines the trajectory and the tactical level executes the high-level decisions. This high-level decision system is the main goal of this research. To address realistic experiments, we set up three scenarios: intersections with traffic lights, intersections with traffic signs and uncontrolled intersections. The results of this paper show that a Proximal Policy Optimization algorithm can infer ego vehicle-desired behavior for different intersection scenarios based only on the behavior of adversarial vehicles.

Driver Take-Over Behaviour Study Based on Gaze Focalization and Vehicle Data in CARLA Simulator.

Autonomous vehicles are the near future of the automobile industry. However, until they reach Level 5, humans and cars will share this intermediate future. Therefore, studying the transition between autonomous and manual modes is a fascinating topic. Automated vehicles may still need to occasionally hand the control to drivers due to technology limitations and legal requirements. This paper presents a study of driver behaviour in the transition between autonomous and manual modes using a CARLA simulator. To our knowledge, this is the first take-over study with transitions conducted on this simulator. For this purpose, we obtain driver gaze focalization and fuse it with the road's semantic segmentation to track to where and when the user is paying attention, besides the actuators' reaction-time measurements provided in the literature. To track gaze focalization in a non-intrusive and inexpensive way, we use a method based on a camera developed in previous works. We devised it with the OpenFace 2.0 toolkit and a NARMAX calibration method. It transforms the face parameters extracted by the toolkit into the point where the user is looking on the simulator scene. The study was carried out by different users using our simulator, which is composed of three screens, a steering wheel and pedals. We distributed this proposal in two different computer systems due to the computational cost of the simulator based on the CARLA simulator. The robot operating system (ROS) framework is in charge of the communication of both systems to provide portability and flexibility to the proposal. Results of the transition analysis are provided using state-of-the-art metrics and a novel driver situation-awareness metric for 20 users in two different scenarios.

Gaze Focalization System for Driving Applications Using OpenFace 2.0 Toolkit with NARMAX Algorithm in Accidental Scenarios.

Monitoring driver attention using the gaze estimation is a typical approach used on road scenes. This indicator is of great importance for safe driving, specially on Level 3 and Level 4 automation systems, where the take over request control strategy could be based on the driver's gaze estimation. Nowadays, gaze estimation techniques used in the state-of-the-art are intrusive and costly, and these two aspects are limiting the usage of these techniques on real vehicles. To test this kind of application, there are some databases focused on critical situations in simulation, but they do not show real accidents because of the complexity and the danger to record them. Within this context, this paper presents a low-cost and non-intrusive camera-based gaze mapping system integrating the open-source state-of-the-art OpenFace 2.0 Toolkit to visualize the driver focalization on a database composed of recorded real traffic scenes through a heat map using NARMAX (Nonlinear AutoRegressive Moving Average model with eXogenous inputs) to establish the correspondence between the OpenFace 2.0 parameters and the screen region the user is looking at. This proposal is an improvement of our previous work, which was based on a linear approximation using a projection matrix. The proposal has been validated using the recent and challenging public database DADA2000, which has 2000 video sequences with annotated driving scenarios based on real accidents. We compare our proposal with our previous one and with an expensive desktop-mounted eye-tracker, obtaining on par results. We proved that this method can be used to record driver attention databases.

Drive-By-Wire Development Process Based on ROS for an Autonomous Electric Vehicle.

This paper presents the development process of a robust and ROS-based Drive-By-Wire system designed for an autonomous electric vehicle from scratch over an open source chassis. A revision of the vehicle characteristics and the different modules of our navigation architecture is carried out to put in context our Drive-by-Wire system. The system is composed of a Steer-By-Wire module and a Throttle-By-Wire module that allow driving the vehicle by using some commands of lineal speed and curvature, which are sent through a local network from the control unit of the vehicle. Additionally, a Manual/Automatic switching system has been implemented, which allows the driver to activate the autonomous driving and safely taking control of the vehicle at any time. Finally, some validation tests were performed for our Drive-By-Wire system, as a part of our whole autonomous navigation architecture, showing the good working of our proposal. The results prove that the Drive-By-Wire system has the behaviour and necessary requirements to automate an electric vehicle. In addition, after 812 h of testing, it was proven that it is a robust Drive-By-Wire system, with high reliability. The developed system is the basis for the validation and implementation of new autonomous navigation techniques developed within the group in a real vehicle.

A Waypoint Tracking Controller for Autonomous Road Vehicles Using ROS Framework.

Automated Driving Systems (ADSs) require robust and scalable control systems in order to achieve a safe, efficient and comfortable driving experience. Most global planners for autonomous vehicles provide as output a sequence of waypoints to be followed. This paper proposes a modular and scalable waypoint tracking controller for Robot Operating System (ROS)-based autonomous guided vehicles. The proposed controller performs a smooth interpolation of the waypoints and uses optimal control techniques to ensure robust trajectory tracking even at high speeds in urban environments (up to 50 km/h). The delays in the localization system and actuators are compensated in the control loop to stabilize the system. Forward velocity is adapted to path characteristics using a velocity profiler. The controller has been implemented as an ROS package providing scalability and exportability to the system in order to be used with a wide variety of simulators and real vehicles. We show the results of this controller using the novel and hyper realistic CARLA Simulator and carrying out a comparison with other standard and state-of-art trajectory tracking controllers.

Robustifying semantic cognition of traversability across wearable RGB-depth cameras.

Semantic segmentation represents a promising means to unify different detection tasks, especially pixel-wise traversability perception as the fundamental enabler in robotic vision systems aiding upper-level navigational applications. However, major research efforts are being put into earning marginal accuracy increments on semantic segmentation benchmarks, without assuring the robustness of real-time segmenters to be deployed in assistive cognition systems for the visually impaired. In this paper, we explore in a comparative study across four perception systems, including a pair of commercial smart glasses, a customized wearable prototype, and two portable red-green-blue-depth (RGB-D) cameras that are being integrated in the next generation of navigation assistance devices. More concretely, we analyze the gap between the concepts of "accuracy" and "robustness" on the critical traversability-related semantic scene understanding. A cluster of efficient deep architectures is proposed, which is built using spatial factorizations, hierarchical dilations, and pyramidal representations. Based on these architectures, this research demonstrates the augmented robustness of semantically traversable area parsing against the variations of environmental conditions in diverse RGB-D observations, and sensorial factors such as illumination, imaging quality, field of view, and detectable depth range.

Real-Time Semantic Segmentation for Fisheye Urban Driving Images Based on ERFNet.

The interest in fisheye cameras has recently risen in the autonomous vehicles field, as they are able to reduce the complexity of perception systems while improving the management of dangerous driving situations. However, the strong distortion inherent to these cameras makes the usage of conventional computer vision algorithms difficult and has prevented the development of these devices. This paper presents a methodology that provides real-time semantic segmentation on fisheye cameras leveraging only synthetic images. Furthermore, we propose some Convolutional Neural Networks(CNN) architectures based on Efficient Residual Factorized Network(ERFNet) that demonstrate notable skills handling distortion and a new training strategy that improves the segmentation on the image borders. Our proposals are compared to similar state-of-the-art works showing an outstanding performance and tested in an unknown real world scenario using a fisheye camera integrated in an open-source autonomous electric car, showing a high domain adaptation capability.

Unifying Terrain Awareness for the Visually Impaired through Real-Time Semantic Segmentation.

Navigational assistance aims to help visually-impaired people to ambulate the environment safely and independently. This topic becomes challenging as it requires detecting a wide variety of scenes to provide higher level assistive awareness. Vision-based technologies with monocular detectors or depth sensors have sprung up within several years of research. These separate approaches have achieved remarkable results with relatively low processing time and have improved the mobility of impaired people to a large extent. However, running all detectors jointly increases the latency and burdens the computational resources. In this paper, we put forward seizing pixel-wise semantic segmentation to cover navigation-related perception needs in a unified way. This is critical not only for the terrain awareness regarding traversable areas, sidewalks, stairs and water hazards, but also for the avoidance of short-range obstacles, fast-approaching pedestrians and vehicles. The core of our unification proposal is a deep architecture, aimed at attaining efficient semantic understanding. We have integrated the approach in a wearable navigation system by incorporating robust depth segmentation. A comprehensive set of experiments prove the qualified accuracy over state-of-the-art methods while maintaining real-time speed. We also present a closed-loop field test involving real visually-impaired users, demonstrating the effectivity and versatility of the assistive framework.

Model-Based Real-Time Non-Rigid Tracking.

This paper presents a sequential non-rigid reconstruction method that recovers the 3D shape and the camera pose of a deforming object from a video sequence and a previous shape model of the object. We take PTAM (Parallel Mapping and Tracking), a state-of-the-art sequential real-time SfM (Structure-from-Motion) engine, and we upgrade it to solve non-rigid reconstruction. Our method provides a good trade-off between processing time and reconstruction error without the need for specific processing hardware, such as GPUs. We improve the original PTAM matching by using descriptor-based features, as well as smoothness priors to better constrain the 3D error. This paper works with perspective projection and deals with outliers and missing data. We evaluate the tracking algorithm performance through different tests over several datasets of non-rigid deforming objects. Our method achieves state-of-the-art accuracy and can be used as a real-time method suitable for being embedded in portable devices.

A Multi-Sensorial Simultaneous Localization and Mapping (SLAM) System for Low-Cost Micro Aerial Vehicles in GPS-Denied Environments.

One of the main challenges of aerial robots navigation in indoor or GPS-denied environments is position estimation using only the available onboard sensors. This paper presents a Simultaneous Localization and Mapping (SLAM) system that remotely calculates the pose and environment map of different low-cost commercial aerial platforms, whose onboard computing capacity is usually limited. The proposed system adapts to the sensory configuration of the aerial robot, by integrating different state-of-the art SLAM methods based on vision, laser and/or inertial measurements using an Extended Kalman Filter (EKF). To do this, a minimum onboard sensory configuration is supposed, consisting of a monocular camera, an Inertial Measurement Unit (IMU) and an altimeter. It allows to improve the results of well-known monocular visual SLAM methods (LSD-SLAM and ORB-SLAM are tested and compared in this work) by solving scale ambiguity and providing additional information to the EKF. When payload and computational capabilities permit, a 2D laser sensor can be easily incorporated to the SLAM system, obtaining a local 2.5D map and a footprint estimation of the robot position that improves the 6D pose estimation through the EKF. We present some experimental results with two different commercial platforms, and validate the system by applying it to their position control.

Visual Object Recognition with 3D-Aware Features in KITTI Urban Scenes.

Driver assistance systems and autonomous robotics rely on the deployment of several sensors for environment perception. Compared to LiDAR systems, the inexpensive vision sensors can capture the 3D scene as perceived by a driver in terms of appearance and depth cues. Indeed, providing 3D image understanding capabilities to vehicles is an essential target in order to infer scene semantics in urban environments. One of the challenges that arises from the navigation task in naturalistic urban scenarios is the detection of road participants (e.g., cyclists, pedestrians and vehicles). In this regard, this paper tackles the detection and orientation estimation of cars, pedestrians and cyclists, employing the challenging and naturalistic KITTI images. This work proposes 3D-aware features computed from stereo color images in order to capture the appearance and depth peculiarities of the objects in road scenes. The successful part-based object detector, known as DPM, is extended to learn richer models from the 2.5D data (color and disparity), while also carrying out a detailed analysis of the training pipeline. A large set of experiments evaluate the proposals, and the best performing approach is ranked on the KITTI website. Indeed, this is the first work that reports results with stereo data for the KITTI object challenge, achieving increased detection ratios for the classes car and cyclist compared to a baseline DPM.

Complete low-cost implementation of a teleoperated control system for a humanoid robot.

Humanoid robotics is a field of a great research interest nowadays. This work implements a low-cost teleoperated system to control a humanoid robot, as a first step for further development and study of human motion and walking. A human suit is built, consisting of 8 sensors, 6 resistive linear potentiometers on the lower extremities and 2 digital accelerometers for the arms. The goal is to replicate the suit movements in a small humanoid robot. The data from the sensors is wirelessly transmitted via two ZigBee RF configurable modules installed on each device: the robot and the suit. Replicating the suit movements requires a robot stability control module to prevent falling down while executing different actions involving knees flexion. This is carried out via a feedback control system with an accelerometer placed on the robot's back. The measurement from this sensor is filtered using Kalman. In addition, a two input fuzzy algorithm controlling five servo motors regulates the robot balance. The humanoid robot is controlled by a medium capacity processor and a low computational cost is achieved for executing the different algorithms. Both hardware and software of the system are based on open platforms. The successful experiments carried out validate the implementation of the proposed teleoperated system.

Assisting the visually impaired: obstacle detection and warning system by acoustic feedback.

The aim of this article is focused on the design of an obstacle detection system for assisting visually impaired people. A dense disparity map is computed from the images of a stereo camera carried by the user. By using the dense disparity map, potential obstacles can be detected in 3D in indoor and outdoor scenarios. A ground plane estimation algorithm based on RANSAC plus filtering techniques allows the robust detection of the ground in every frame. A polar grid representation is proposed to account for the potential obstacles in the scene. The design is completed with acoustic feedback to assist visually impaired users while approaching obstacles. Beep sounds with different frequencies and repetitions inform the user about the presence of obstacles. Audio bone conducting technology is employed to play these sounds without interrupting the visually impaired user from hearing other important sounds from its local environment. A user study participated by four visually impaired volunteers supports the proposed system.

Non-linearity analysis of depth and angular indexes for optimal stereo SLAM.

In this article, we present a real-time 6DoF egomotion estimation system for indoor environments using a wide-angle stereo camera as the only sensor. The stereo camera is carried in hand by a person walking at normal walking speeds 3-5 km/h. We present the basis for a vision-based system that would assist the navigation of the visually impaired by either providing information about their current position and orientation or guiding them to their destination through different sensing modalities. Our sensor combines two different types of feature parametrization: inverse depth and 3D in order to provide orientation and depth information at the same time. Natural landmarks are extracted from the image and are stored as 3D or inverse depth points, depending on a depth threshold. This depth threshold is used for switching between both parametrizations and it is computed by means of a non-linearity analysis of the stereo sensor. Main steps of our system approach are presented as well as an analysis about the optimal way to calculate the depth threshold. At the moment each landmark is initialized, the normal of the patch surface is computed using the information of the stereo pair. In order to improve long-term tracking, a patch warping is done considering the normal vector information. Some experimental results under indoor environments and conclusions are presented.

Error analysis in a stereo vision-based pedestrian detection sensor for collision avoidance applications.

This paper presents an analytical study of the depth estimation error of a stereo vision-based pedestrian detection sensor for automotive applications such as pedestrian collision avoidance and/or mitigation. The sensor comprises two synchronized and calibrated low-cost cameras. Pedestrians are detected by combining a 3D clustering method with Support Vector Machine-based (SVM) classification. The influence of the sensor parameters in the stereo quantization errors is analyzed in detail providing a point of reference for choosing the sensor setup according to the application requirements. The sensor is then validated in real experiments. Collision avoidance maneuvers by steering are carried out by manual driving. A real time kinematic differential global positioning system (RTK-DGPS) is used to provide ground truth data corresponding to both the pedestrian and the host vehicle locations. The performed field test provided encouraging results and proved the validity of the proposed sensor for being used in the automotive sector towards applications such as autonomous pedestrian collision avoidance.