Loading...
2 results
Search Results
Now showing 1 - 2 of 2
- Adaptive Convolutional Neural Network for Predicting Steering Angle and Acceleration on Autonomous Driving ScenarioPublication . Vasiljević, Ive; Musić, Josip; Mendes, João; Lima, JoséThis paper introduces a novel approach to autonomous vehicle control using an end-to-end learning framework. While existing solutions in the field often rely on computationally expensive architectures, our proposed lightweight model achieves comparable efficiency. We leveraged the Car Learning to Act (CARLA) simulator to generate training data by recording sensor inputs and corresponding control actions during simulated driving. The Mean Squared Error (MSE) loss function served as a performance metric during model training. Our end-to-end learning architecture demonstrates promising results in predicting steering angle and throttle, offering a practical and accessible solution for autonomous driving. Results of the experiment showed that our proposed network is ≈ 5.4 times lighter than Nvidia’s PilotNet and had a slightly lower testing loss. We showed that our network is offering a balance between performance and computational efficiency. By eliminating the need for handcrafted feature engineering, our approach simplifies the control process and reduces computational demands. Experimental evaluation on a testing map showcases the model’s effectiveness in real-world scenarios whilst being competitive with other existing models.
- Object detection for indoor localization systemPublication . Braun, João; Mendes, João; Pereira, Ana I.; Lima, José; Costa, Paulo Gomes daThe urge for robust and reliable localization systems for autonomous mobile robots (AMR) is increasing since the demand for these automated systems is rising in service, industry, and other areas of the economy. The localization of AMRs is one of the crucial challenges, and several approaches exist to solve this. The most well-known localization systems are based on LiDAR data due to their reliability, accuracy, and robustness. One standard method is to match the reference map information with the actual readings from LiDAR or camera sensors, allowing localization to be performed. However, this approach has difficulties handling anything that does not belong to the original map since it affects the matching algorithm’s performance. Therefore, they should be considered outliers. In this paper, a deep learning-based object detection algorithm is not only used for detection but also to classify them as outliers from the localization’s perspective. This is an innovative approach to improve the localization results in a real mobile platform. Results are encouraging, and the proposed methodology is being tested in a real robot.