Multi-sensor 3D object detection is essential for autonomous driving; however, it faces three persistent challenges: scale variations across objects, structural degradation of fine-grained features during view transformation, and ambiguous object boundaries in complex scenes. To address these issues, we propose FGA-BEVFusion, a LiDAR–camera bird's-eye-view (BEV) fusion framework built upon a novel frequency-guided attention (FGA) mechanism that explicitly preserves structural information in the frequency domain. FGA operates in the view space before BEV projection: a Multi-Spectral Channel Attention (MCA) module adaptively reweights spectral responses, and a Structure-Aware Edge Injection (SEI) module enhances spatial boundary sensitivity. In the BEV space, a Sparsity-Aware BEV Refinement (SAR) module performs occupancy-guided hierarchical fusion and spatial recalibration to reduce information loss during decoding. Experiments on the nuScenes benchmark show that FGA-BEVFusion achieves 70.8% mAP and 72.3% NDS on the validation set, with consistent improvements on small and distant object categories and under challenging conditions such as rain and nighttime.
- Article type
- Year
- Co-author
Open Access
Research Article
Just Accepted
Open Access
Research Article
Issue
Accurate trajectory forecasting is essential for enabling autonomous vehicles to navigate safely in complex traffic environments. Current models typically assume that predictions are latency-free, an idealistic simplification that fails in real-world settings, where processing, computation, and transmission inevitably introduce delays. During this latency window, target agents continue to move, which renders early predictions obsolete, thus degrading system performance. To address this challenge, we introduce latency-aware trajectory prediction, a new task that explicitly accounts for latency and repurposes it as a useful signal. We present LatenAux, a consolidated auxiliary learning paradigm that first decouples prediction into two tasks: a primary task that predicts valid-horizon trajectories from historical data, and an auxiliary task that utilizes latency-inclusive observations. By allowing the auxiliary branch access to latency-crafted inputs, LatenAux is then committed to transferring latency-aware knowledge to the primary branch via a progressive feature alignment strategy. This enables the primary model to internalize latency cues without explicit reliance on latency data. Our method departs from conventional auxiliary learning by introducing a soft feature-consistency function, which gradually incorporates auxiliary representations across both scene context and query state levels, enriching features while avoiding overconstraint. In addition, auxiliary queries act as informative priors for the primary branch to further enhance prediction accuracy. Extensive experiments on two large-scale real-world datasets demonstrate the effectiveness and superiority of LatenAux, showing that it consistently supports latency-aware modeling and delivers more accurate and reliable trajectory forecasts.
京公网安备11010802044758号