STFNET: Sparse Temporal Fusion for 3D Object Detection in LiDAR Point Cloud

Abstract

In autonomous driving and robotics, 3D object detection using LiDAR point clouds is a critical task. However, existing single-frame 3D object detection methods face challenges such as noise, occlusions, and sparsity, which degrade detection performance. To address these, we propose the sparse temporal fusion network (STFNet), which leverages multiframe historical information to improve 3D object detection accuracy. The contribution of STFNet contains three core modules: multihistory feature alignment module (MFAM), sparse feature extraction module (SFEM), and temporal fusion transformer (TFformer). MFAM: Ego-motion is used for compensation to align frames, establishing correlations between adjacent frames along the temporal dimension. SFEM: Sparse extraction is performed on features from different time steps to obtain key features within the time series. TFformer: The advanced temporal fusion attention mechanism is introduced to facilitate deep interactions between the current and historical frames. We validated the effectiveness of STFNet on the nuScenes dataset, achieving 71.8% NuScenes detection score (NDS) and 67.0% mean average precision (mAP). Compared to the benchmark method, our method improves 1.6% NDS and 1.5% mAP. Extensive experiments demonstrate that STFNet significantly outperforms most existing methods, highlighting the superiority and generalizability of our approach.