The abilities to understand the social interaction behaviors between a vehicle and its surroundings while predicting its trajectory in an urban environment are critical for road safety in autonomous driving. Social interactions are hard to explain because of their uncertainty. In recent years, neural network-based methods have been widely used for trajectory prediction and have been shown to outperform hand-crafted methods. However, these methods suffer from their lack of interpretability. In order to overcome this limitation, we combine the interpretability of a discrete choice model with the high accuracy of a neural network-based model for the task of vehicle trajectory prediction in an interactive environment. We implement and evaluate our model using the INTERACTION dataset and demonstrate the effectiveness of our proposed architecture to explain its predictions without compromising the accuracy.

We introduce a new approach for multiscale 3Dsemantic scene completion from voxelized sparse 3D LiDAR scans. As opposed to the literature, we use a 2D UNet backbone with comprehensive multiscale skip connections to enhance feature flow, along with 3D segmentation heads. On the SemanticKITTI benchmark, our method performs on par on semantic completion and better on occupancy completion than all other published methods - while being significantly lighter and faster. As suchit provides a great performance/speed trade-off for mobile-robotics applications. The ablation studies demonstrate our method is robust to lower density inputs, and that it enables very high speed semantic completion at the coarsest level. Our code is available at https://github.com/cv-rits/LMSCNet.

Autonomous driving in urban environment needs to anticipate a number of dangerous events, such as the presence of moving vehicles in occluded areas. This paper presents an approach computing a safe motion along a fixed path in an urban environment with dynamic vehicles that may be occluded. The method works on the time-path space and uses a visibility graph to compute the speed profile, considering both safety and comfort. Evaluations performed on CARLA simulator on several typical scenarios show that the approach is able to drive safely in presence of hidden obstacles.

For an efficient integration of autonomous vehicles on roads, human-like reasoning and decision making in complex traffic situations are needed. One of the key factors to achieve this goal is the estimation of the future behavior of the vehicles present in the scene. In this work, we propose a new approach to predict the motion of vehicles surrounding a target vehicle in a highway environment. Our approach is based on an LSTM encoder-decoder that uses a social pooling mechanism to model the interactions between all the neighboring vehicles. The originality of our social pooling module is that it combines both local and non-local operations. The non-local multi-head attention mechanism captures the relative importance of each vehicle despite the inter-vehicle distances to the target vehicle, while the local blocks represent nearby interactions between vehicles. This paper compares the proposed approach with the state-of-the-art using two naturalistic driving datasets: Next Generation Simulation (NGSIM) and the new highD Dataset. The proposed method outperforms existing ones in terms of RMS values of prediction error, which shows the effectiveness of combining local and non-local operations in such a context.

We present here a framework to integrate into a motion planning method the interaction zones of a moving robot with its future surroundings, the reachable interaction sets. It can handle highly dynamic scenarios when combined with path planning methods optimized for quasi-static environments. As a demonstrator, it is integrated here with an artificial potential field reactive method and with a Bézler curve path planning. Experimental evaluations show that this approach significantly improves dynamic path planning methods, especially when the speeds of the obstacles are higher than the one of the robot. The presented approach is used together with a global planning approach in order to handle complex static environments in presence of fast-moving obstacles. When the ego vehicle is not holonomic the presented approach is able to take dynamic constraints into account, which improve the prediction accuracy.

Simultaneous Localization And Mapping (SLAM) is one of the major bricks needed to build truly autonomous mobile robots. The probabilistic formulation of SLAM is based on two models: the motion model and the observation model. In practice, these models, together with the SLAM map representation, do not model perfectly the robot's real dynamics, the sensor measurement errors and the environment. Consequently, systematic errors affect SLAM estimations. In this paper, we propose two approaches to predict corrections to be applied to SLAM estimations. Both are based on the Ensemble Multilayer Perceptron model. The first approach uses successive estimated poses to predict the errors, with no assumptions on the underlying SLAM process or sensor used. The second method is specific to 2D likelihood SLAM approaches, thus, the likelihood distributions are used to predict the corrections, making this second approach independent of the sensor used. We also build a hybrid correction module based on successive estimated poses and the likelihood distributions. The validity of both approaches is evaluated through two experiments using different evaluation metrics and sensor configurations.

Learning from human driver's strategies for solving complex and potentially dangerous situations including interaction with other road users has the potential to improve decision-making methods for automated vehicles. In this paper, we focus on simple unsignalized intersections and roundabouts in presence of another vehicle. We propose a human-like decision-making algorithm for these scenarios built up from human drivers recordings. The algorithm includes a risk assessment to avoid collisions in the intersection area. Three road topologies with different interaction scenarios were presented to human participants on a previously developed simulation tool. The same scenarios have been used to validate our decision-making process. The algorithm showed promising results with no collisions in all setups and the ability to successfully determine to go before or after another vehicle.

In this paper, we propose an automatic approach for plant species identification, based on the visual information provided by the plant leaves. More precisely, we consider two sources of information: the leaf margin and the leaf salient points. We investigate two shape context based descriptors: the first one describes the leaf boundary while the second descriptor represents the spatial correlation between salient points of the leaf and its margin. We also study the performance of the fusion of these two descriptors on the Image-CLEF 2011 and 2012 leaf datasets. Experiments show the effectiveness and the efficiency of the proposed method.

In this paper we introduce a new multiscale shape-based approach for leaf image retrieval. The leaf is represented by local descriptors associated with margin sample points. Within this local description, we study four multiscale triangle representations: the well known triangle area representation (TAR), the triangle side lengths representation (TSL) and two new representations that we denote triangle oriented angles (TOA) and triangle side lengths and angle representation (TSLA). Unlike existing TAR approaches, where a global matching is performed, the similarity measure is based on a locality sensitive hashing of local descriptors. The proposed approach is invariant under translation, rotation and scale and robust under partial occlusion. Evaluations made on four public leaf datasets show that our shape-based approach achieves a high retrieval accuracy w.r.t. state-of-art methods.

This paper presents an Android application for plant identification. The system relies on the observation of leaf images. Unlike other mobile plant identification applications, the user may choose the leaf characters that will guide the identification process. For this purpose, two kinds of descriptors are proposed to the user: a shape descriptor based on a multiscale triangular representation of the leaf margin and a descriptor of the salient points of the leaf. The application achieves good identification accuracy and provides Android users a useful system for plant identification.

This paper presents a novel method for leaf species identification combining local and shape-based features. Our approach extends the shape context model in two ways. First of all, two different sets of points are distinguished when computing the shape contexts: the voting set, i.e. the points used to describe the coarse arrangement of the shape and the computing set containing the points where the shape contexts are computed. This representation is enriched by introducing local features computed in the neighborhood of the computing points. Experiments show the effectiveness of our approach.

In this paper, we report on work around the I-SEARCH EU (FP7 ICT STREP) project whose objective is the development of a multimodal search engine targeted at mobile and desktop devices. Each of these device classes has its specific hardware capabilities and set of supported features. In order to provide a common multimodal search experience across device classes, one size does not fit all. We highlight ways to achieve the same functionality agnostic of the device class being used for the search, and present concrete use cases.

This paper presents a new approach to 3D shape comparison and retrieval based on the computation of Gaussian transforms of the surface model on a set of points regularly distributed inside the model's bounding box. A study of the properties of the Gaussian term leads us to an efficient computation of the description. This method is evaluated on the Princeton Shape Benchmark database.

In this paper we present a new method for alignment of 3D models. This approach is based on symmetry properties, and uses the fact that the principal components analysis (PCA) have good properties with respect to the planar reflective symmetry. The fast search of the best optimal alignment axes within the PCA-eigenvectors is an essential first step in our alignment process. The plane reflection symmetry is used as a criterion for selection. This pre-processing transforms the alignment problem into an indexing scheme based on the number of the retained PCA-axes. We also introduce a local translational invariance cost (LTIC) that captures a measure of the local translational symmetries of a shape with respect to a given direction. Experimental results show that the proposed method finds the rotation that best aligns a 3D mesh.

We present here a new descriptor for depth images adapted to 2D/3D model matching and retrieving. We propose a representation of a 3D model by 20 depth images rendered from the vertices of a regular dodecahedron. One depth image of a 3D model is associated to a set of depth lines which will be afterward transformed into sequences. The depth sequence information provides a more accurate description of 3D shape boundaries than using other 2D shape descriptors. Similarity computing is performed when dynamic programming distance (DPD) is used to compare the depth line descriptors. The DPD leads to an accurate matching of sequences even in the presence of local shifting on the shape. Results on a large 3D database show efficiency of our 2D/3D approach.

In this paper, we propose a novel 2D/3D approach for 3D model matching and retrieving. Each model is represented by a set of depth lines which will be afterward transformed into sequences. The depth sequence information provides a more accurate description of 3D shape boundaries than using other 2D shape descriptors. Retrieval is performed when dynamic programming distance (DPD) is used to compare the depth line descriptors. The DPD leads to an accurate matching of sequences even in the presence of local shifting on the shape. Experimentally, we show absolute improvement in retrieval performance on the Princeton 3D Shape Benchmark database.

We present a new approach for 3D model indexing and retrieval using 2D/3D shape descriptors based on silhouettes or depth-buffer images. To take into account the dispersion of information in the views, we associate to each view a relevance index which will be afterward used in the dissimilarity computation. The performance of this new approach is evaluated on the Princeton 3D shape benchmark database

We propose a method to model and control topological changes by a smooth deformation of a polyhedral mesh using curves and loops. As changing the genus of a surface is not a continuous transformation, the topological change is made when an intermediate shape between the two topologies has been obtained. The creation and the deletion of holes are studied. The deletion of a hole uses non null-homotopic loops to designate the hole to be deleted. A method computing two independent loops associated to a hole is presented.

A closed orientable surface of genus $g$ can be obtained by appropriat e identification of pairs of edges of a $4g$-gon (the polygonal schema). The identified edges form $2g$ loops on the surface, that are disjoint except for their common end-point. These loops are generators of both the fundamental group and the homology group of the surface. The inverse problem is concerned with finding a set of $2g$ loops on a triangulated surface, such that cutting the surface along these loops yields a (canonical) polygonal schema. We present two optimal algorithms for this inverse problem. Both algorithms have been implemented using the CGAL polyhedron data structure.

We introduce a method for extracting skeletal curves from an unorganized collection of scattered data points lying on a surface. These curves may have a tree like structure to capture branching shapes such as blood vessels. The skeletal curves can be used for different applications ranging from surface reconstruction to object recognition.

A new technique is presented for computing continuous shape transformations between polyhedral objects. The correspondence and the interpolation problems are considered jointly as we construct and process on sampling polyhedral meshes of the objects. The approach is adapted to objects that are star-shaped around an axis. The process gives the animator a high level control over the shape transformation by providing natural specification and interaction. © 1997 John Wiley & Sons, Ltd.

ISBN: 0-201-58889-7

Abstract We propose an integrated set of methods for designing and animating bodies whose deformable flesh coats an articulated skeleton. The proposed methods provide good automatic positioning of the ‘skin’ after movements, and automatic deformation after collisions with rigid or deformable objects (dynamic or static). The response to collisions includes a feedback from the flesh to the skeleton, whose movement is adequately modified. Moreover, coating the skeleton gives an easy solution to closed loop collisions and angle constraints control. Our model is modular and gives some high level control to the user.