Skip to Main content Skip to Navigation

Fusion Caméra-LiDAR à l’aide de réseaux de neurones artificiels pour la détection d’obstacles à bord de véhicules autonomes

Abstract : Boosted by technological advances on sensors, especially on LiDARsensors, and deep learningtechniques, object detection on 3D scenes has become a major stake for the deployment ofautonomous vehicles. However, most of the state-of-the-art methods tend to produce more errorswhen they exploit data from sensors that were not used for the training stage.In this PhD thesis, we investigate solutions to improve the portability of detectors acrossvarious LiDAR models. In fact, a generalizable detector can be easily embedded in new vehicles.Moreover, such a detector can ease the annotation process by providing an initial set of labels onrecorded data instead of annotating from scratch.We first introduce the context of the thesis by defining the studied tasks aswell as the datasetsand the metrics that are used in this work.Thereafter, we trace the evolution of deep learning techniques for object detection. Technologicaladvances on RGB images heavily influenced the other fields. Hence, a history of 2Ddetection methods are introduced first and foremost. Then, a listing of state-of-the-art 3D objectdetection methods is developed. The topic has been greatly investigated. Consequently, thesemethods are summarized and grouped according to the sensor types they use.The following chapter presents our first contribution, an anchor-free LiDAR detector in orderto reduce the number of hyperparameters required for the training. The network is a segmentationnetwork whose goal is to produce a pseudo occupancy map representing a top view of the scene.Each pixel of the map indicate an estimation of the occupancy of its corresponding 3D region.Obstacles are represented as ellipses on the map to allow the extraction of their parametersthrough image processing and optimization techniques.Lastly, our second contribution deals with portability of detectors across datasets built withvarious setups. The goal is to determine how to improve the performances of a detector whentrained on one dataset and applied on others without supplementary priors. This study revolvesaround one LiDAR detector and two camera - LiDAR networks. In order to improve the resistanceto domain variations, two techniques were developed. The first one consists in the randomizedremoval of layers in the input point cloud. This technique allows to increase the variability of dataobserved during the training stage. The second technique consists in representing the obstaclesas Gaussian functions, Statistical distances then can be used as loss functions. Both techniquesimprove the performances, especially on low-resolutions point clouds. Moreover, the LiDARdetector, besides running in real-time, is able to provide reliable detections on multiple datasets.
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Monday, June 20, 2022 - 9:05:12 AM
Last modification on : Wednesday, June 22, 2022 - 3:06:30 AM


Version validated by the jury (STAR)


  • HAL Id : tel-03699119, version 1


Ruddy Théodose. Fusion Caméra-LiDAR à l’aide de réseaux de neurones artificiels pour la détection d’obstacles à bord de véhicules autonomes. Vision par ordinateur et reconnaissance de formes [cs.CV]. Université Clermont Auvergne, 2021. Français. ⟨NNT : 2021UCFAC090⟩. ⟨tel-03699119⟩



Record views


Files downloads