Research

BEVFlow: Streamlined Excellence in Instance Prediction from a Bird’s-Eye View

The BEVFlow architecture addresses the challenge of accurate motion prediction in autonomous vehicles by integrating centripetal motion prediction with backward flow warping. It comprises three main components: a feature mapping and camera fusion module (built on LSS) that converts multi-view camera inputs into BEV feature maps over temporal frames; a segmentation and flow prediction module that leverages spatio-temporal data to generate segmentation maps and centripetal backward flows for future frames; and a warping-based post-processing stage to reconstruct instance predictions. Optimization efforts focus on enhancing accuracy and efficiency through hyperparameter tuning, such as batch size and learning rate adjustments.

Affordance Learning for Detection of Drivable Area

Research

Affordance learning for drivable areas focuses on leveraging deep learning techniques to predict actionable cues that determine navigable regions for autonomous vehicles. This involves training neural networks on multi-modal sensor data, such as LiDAR, camera imagery, and radar, to map low-level perceptual inputs to high-level affordances like lane boundaries, free space, and obstacle proximity. State-of-the-art models integrate spatial and temporal features using convolutional and recurrent architectures to enhance real-time decision-making. The learning process often employs supervised or self-supervised methods, emphasizing generalization across diverse road environments while maintaining robustness to occlusions and dynamic objects.

Research
Research

Localization and Mapping(SLAM)

Research

ReConfNET, an advanced confidence-driven RGB-D SLAM framework tailored for high-precision indoor scene reconstruction. The system leverages a hierarchical generative model integrated with a confidence-based loss function to optimize pose estimation and scene representation. By processing continuous RGB-D input streams, ReConfNET extracts high-quality key points and incorporates neural implicit encoding combined with volume rendering techniques to generate multi-resolution, scalable scene reconstructions. Its architecture ensures accurate spatial mapping by dynamically adjusting confidence metrics to refine depth and pose predictions. Comprehensive evaluations reveal ReConfNET’s superiority in achieving higher reconstruction fidelity and computational efficiency over traditional SLAM methods, making it well-suited for applications in robotics, augmented reality (AR), and intelligent environments.

close
close
close
mail

info@kletech.ac.in

coe@kletech.ac.in (Controller of Examinations)