https://blog.pouceheure.fr/tags/ai_ml/Recent content in Ai_ml on Hugo POUSSEURHugoenThu, 28 Mar 2024 00:00:00 +0000https://blog.pouceheure.fr/articles/thesis/Thu, 28 Mar 2024 00:00:00 +0000https://blog.pouceheure.fr/articles/thesis/Abstract My thesis is based on shared navigation between the autonomous system and the human. In our research, we focus on command fusion. In our approach, both entities, the human and the autonomous system, simultaneously control the vehicle, and a module acquires their commands and performs the command fusion. This approach involves studying the intentions of both the human and the autonomous system to ensure the most appropriate fusion of their choices and to evaluate the decision-making of each entity.https://blog.pouceheure.fr/projects/visual-control/Wed, 01 Feb 2023 00:00:00 +0000https://blog.pouceheure.fr/projects/visual-control/Abstract As part of a research project, I contributed to the development and testing of a Visual Servoing framework, a control strategy that links image feature variations to robot velocities. Applied to autonomous vehicles, this approach enables lane-keeping purely from camera input, without GPS or map dependency. It integrates lane detection using deep learning and transforms the output into features used in a control law that centers the vehicle in the lane.https://blog.pouceheure.fr/articles/machine-learning-controller/Sun, 01 Jan 2023 00:00:00 +0000https://blog.pouceheure.fr/articles/machine-learning-controller/Abstract Researchers have been attempting to make the car drive autonomously. The environment perception together with safe guidance and control is an important task and are one of the big challenges when developing this kind of system. Geometrical or physical based models, machine learning based models and those based on a mixture of both models, are the three types of navigation methods used to resolve this problem. The last method takes advantage of the learning capability of machine learning models and uses the safeness of geometric models in order to better perform the navigation task.https://blog.pouceheure.fr/projects/human-prediction/Thu, 10 Mar 2022 00:00:00 +0000https://blog.pouceheure.fr/projects/human-prediction/Abstract A multimodal deep learning model was developed to predict human driving behavior over a short time horizon. The training dataset was recorded specifically for this project. Video 1 - Video demo, human driving behavior prediction. More details about this work are provided in PredictionPousseur 2022 . Introduction Driving intention is represented as a sequence of vehicle states. Let $I$ be: $$ I = \{x_0, x_1, ..., x_n\} \quad \text{Eq (1)} $$ Here, $(x_i) = (v_i, w_i)$ denotes the state at time $i$, where $v_i$ is the linear velocity and $w_i$ is the angular velocity.https://blog.pouceheure.fr/projects/lane-detection/Mon, 10 Jan 2022 00:00:00 +0000https://blog.pouceheure.fr/projects/lane-detection/Abstract Lane detection is an essential function for autonomous vehicles. While GPS and HD maps offer accurate localization under optimal conditions, vision-based detection provides an important backup, especially in areas with weak GNSS signals or outdated maps. This system implements a deep-learning lane detection pipeline using a convolutional autoencoder, designed to identify and segment multiple lane boundaries in various road conditions using only front-facing camera images. The approach integrates image preprocessing, deep-learning inference, optional temporal tracking, and curve fitting to generate stable and clean lane boundaries.https://blog.pouceheure.fr/articles/human-driving-prediction/Sat, 01 Jan 2022 00:00:00 +0000https://blog.pouceheure.fr/articles/human-driving-prediction/Abstract Predicting the intentions of the human and the machine on a near future is required to the human-machine shared control of automated intelligent vehicles. The autonomous system is able to inform about its future intentions, however it is not possible for the human to provide this information, it is necessary then to predict it. This paper proposes a deep learning methodology to predict human navigation intentions in a time horizon of a few seconds, using a recurrent neural network (RNN) architecture based on the Long Short-Term Memory (LSTM) architecture.https://blog.pouceheure.fr/articles/isolated-area-consumption/Fri, 01 Jan 2021 00:00:00 +0000https://blog.pouceheure.fr/articles/isolated-area-consumption/Abstract Forecasting consumption in isolated areas represents a challenging problem typically resolved using deep learning or large mathematical models with various dimensions. These models require expertise in metering and algorithms, and the equipment needs to be frequently maintained. In the context of the MAESHA H2020 project, most of the consumers and producers are isolated. Forecasting becomes more difficult due to the lack of external data and the significant impact of human behaviors on these small systems.https://blog.pouceheure.fr/projects/human-legs-detection/Thu, 10 Dec 2020 00:00:00 +0000https://blog.pouceheure.fr/projects/human-legs-detection/Overview Source Code In addition to the main GitHub project, the following submodules are available: Dataset of scanned legs (with labels): https://github.com/PouceHeure/dataset_lidar2D_legs Labeling GUI tool: https://github.com/PouceHeure/lidar_tool_label Radar interface: https://github.com/PouceHeure/ros_pygame_radar_2D Context This project detects human legs from 2D LiDAR scans using a Recurrent Neural Network (RNN) with LSTM cells. The LSTM processes each scan as a spatial sequence ordered by angle $ \theta $, not as a time series. The model learns local shape patterns in polar space that are typical of legs.https://blog.pouceheure.fr/projects/light-mas/Mon, 07 Sep 2020 00:00:00 +0000https://blog.pouceheure.fr/projects/light-mas/Overview py_light_mas is a lightweight Python framework for multi-agent systems (MAS). It provides simple abstractions for Agents, an Environnemnt, a Simulation loop, and an optional Network for inter-agent communication. Architecture The architecture can be illustrated in two diagrams that describe the same system: Fig. 1 - Relations between Simulation, Environnemnt, Agents, and Network. The Fig. 1 highlights how the Simulation connects the Environnemnt, Agents, and optional Network. Fig. 2 - Event sequence across a simulation tick with message passing.