Rachid Belaroussi’s research while affiliated with Gustave Eiffel University and other places

In this study, an improved deep learning model is proposed to explore the complex interactions between the road environment and driver’s behaviour throughout the generation of a graphical representation. The proposed model consists of an unsupervised Denoising Stacked Autoencoder (SDAE) able to provide output layers in RGB colors. The dataset comes from an experimental driving test where kinematic measures were tracked with an in-vehicle GPS device. The graphical outcomes reveal the method ability to efficiently detect patterns of simple driving behaviors, as well as the road environment complexity and some events encountered along the path.

Fast 3D reconstruction with semantic information in road scenes is of great requirements for autonomous navigation. It involves issues of geometry and appearance in the field of computer vision. In this work, we propose a fast 3D semantic mapping system based on the monocular vision by fusion of localization, mapping, and scene parsing. From visual sequences, it can estimate the camera pose, calculate the depth, predict the semantic segmentation, and finally realize the 3D semantic mapping. Our system consists of three modules: a parallel visual Simultaneous Localization And Mapping (SLAM) and semantic segmentation module, an incrementally semantic transfer from 2D image to 3D point cloud, and a global optimization based on Conditional Random Field (CRF). It is a heuristic approach that improves the accuracy of the 3D semantic labeling in light of the spatial consistency on each step of 3D reconstruction. In our framework, there is no need to make semantic inference on each frame of sequence, since the 3D point cloud data with semantic information is corresponding to sparse reference frames. It saves on the computational cost and allows our mapping system to perform online. We evaluate the system on road scenes, e.g., KITTI, and observe a significant speed-up in the inference stage by labeling on the 3D point cloud.

Fast 3D reconstruction with semantic information on road scenes is of great requirements for autonomous navigation. It involves issues of geometry and appearance in the field of computer vision. In this work, we propose a method of fast 3D semantic mapping based on the monocular vision. At present, due to the inexpensive price and easy installation, monocular cameras are widely equipped on recent vehicles for the advanced driver assistance and it is possible to acquire semantic information and 3D map. The monocular visual sequence is used to estimate the camera pose, calculate the depth, predict the semantic segmentation, and finally realize the 3D semantic mapping by combination of the techniques of localization, mapping and scene parsing. Our method recovers the 3D semantic mapping by incrementally transferring 2D semantic information to 3D point cloud. And a global optimization is explored to improve the accuracy of the semantic mapping in light of the spatial consistency. In our framework, there is no need to make semantic inference on each frame of the sequence, since the mesh data with semantic information is corresponding to sparse reference frames. It saves amounts of the computational cost and allows our mapping system to perform online. We evaluate the system on naturalistic road scenes, e.g., KITTI and observe a significant speed-up in the inference stage by labeling on the mesh.

The bundle of geometry and appearance in computer vision has proven to be a promising solution for robots across a wide variety of applications. Stereo cameras and RGB-D sensors are widely used to realise fast 3D reconstruction and trajectory tracking in a dense way. However, they lack flexibility of seamless switch between different scaled environments, i.e., indoor and outdoor scenes. In addition, semantic information are still hard to acquire in a 3D mapping. We address this challenge by combining the state-of-art deep learning method and semi-dense Simultaneous Localisation and Mapping (SLAM) based on video stream from a monocular camera. In our approach, 2D semantic information are transferred to 3D mapping via correspondence between connective Keyframes with spatial consistency. There is no need to obtain a semantic segmentation for each frame in a sequence, so that it could achieve a reasonable computation time. We evaluate our method on indoor/outdoor datasets and lead to an improvement in the 2D semantic labelling over baseline single frame predictions.

This paper presents a Multi-Hypotheses Tracking (MHT) approach that allows solving ambiguities that arise with previous methods of associating targets and tracks within a highly volatile vehicular environment. The previous approach based on the Dempster-Shafer Theory assumes that associations between tracks and targets are unique; this was shown to allow the formation of ghost tracks when there was too much ambiguity or conflict for the system to take a meaningful decision. The MHT algorithm described in this paper removes this uniqueness condition, allowing the system to include ambiguity and even to prevent making any decision if available data are poor. We provide a general introduction to the Dempster-Shafer Theory and present the previously used approach. Then, we explain our MHT mechanism and provide evidence of its increased performance in reducing the amount of ghost tracks and false positive processed by the tracking system.

18th IEEE International Conference on Intelligent Transportation Systems (ITSC 2015), Las Palmas, Espagne, 15-/09/2015 - 18/09/2015

We are witnessing the clash of two industries and the remaking of in-car market order, as the world of digital knowledge recently made a significant move toward the automotive industry. Mobile operating system providers are battling between each other to take over the in-vehicle entertainment and information systems, while car makers either line up behind their technology or try to keep control over the in-car experience. What is at stake is the map content and location-based services, two key enabling technologies of self-driving cars and future automotive safety systems. These content-based augmented geographic information systems (GIS) as well as Advanced Driver Assistance Systems (ADAS) require an accurate, robust, and reliable estimation of road scene attributes. Accurate localization of the vehicle is a challenging and critical task that natural GPS or classical filter (EKF) cannot reach. This paper proposes a new approach allowing us to give a first answer to the issue of accurate lateral positioning. The proposed approach is based on the fusion of 4 types of data: a GPS, a set of INS/odometer sensors, a road marking detection, and an accurate road marking map. The lateral road markings detection is done with the processing of two lateral cameras and provides an assessment of the lateral distance between the vehicle and the road borders. These information coupled with an accurate digital map of the road markings provide an efficient and reliable way to dramatically improve the localization obtained from only classical way (GPS/INS/Odometer). Moreover, the use of the road marking detection can be done only when the confidence is sufficiently high (punctual use). In fact, the vision processing and the map data can be used punctually only in order to update the classical localization algorithm. The temporary lack of vision data does not affect the quality of lateral positioning. In order to evaluate and validate this approach, a real test scenario was performed on Satory's test track with real embedded sensors. It shows that the lateral estimation of the ego-vehicle positioning is performed with a sub-decimeter accuracy, high enough to be used in autonomous lane keeping, and land-based mobile mapping.

Automated vehicles and Advanced Driver Assistance Systems (ADAS) face a variety of complex situations that are dealt with numerous sensors for the perception of the local driving area. Going forward, we see an increasing use of multiple, different sensors inputs with radar, camera and inertial measurement the most common sensor types. Each system has its own purpose and either displays information or performs an activity without consideration for any other ADAS systems, which does not make the best use of the systems. This paper presents an embedded real-time system to combine the attributes of obstacles, roadway and ego-vehicle features in order to build a collaborative local map. This embedded architecture is called PerSEE: a library of vision-based state-of-the-art algorithms was implemented and distributed in processors of a main fusion electronic board and on smart-cameras board. The embedded hardware architecture of the full PerSEE platform is detailed, with block diagrams to illustrate the partition of the algorithm on the different processors and electronic boards. The communications interfaces as well as the development environment are described.

Reduced visibility on roadways caused by localized fog can impact the traffic flow in many ways: traffic speed, travel time delay, reduced capacity and accident risks. This paper presents a novel approach to estimate visibility conditions using an onboard camera and a digital map. Based on a traffic sign detector's characteristics in the fog, and registering detection by vision and information encoded in the map, we are able to accurately determine the current visual range in hazy conditions. Quantitative results are provided on a large experimental data set of driving environment with various level of fogginess.