TriBand-BEV: Real-Time LiDAR-Only 3D Pedestrian Detection via Height-Aware BEV and High-Resolution Feature Fusion

Abstract: Safe autonomous agents and mobile robots need fast real time 3D perception, especially for vulnerable road users (VRUs) such as pedestrians. We introduce a new bird's eye view (BEV) encoding, which maps the full 3D LiDAR point cloud into a light-weight 2D BEV tensor with three height bands. We explicitly reformulate 3D detection as a 2D detection problem and then reconstruct 3D boxes from the BEV outputs. A single network detects cars, pedestrians, and cyclists in one pass. The backbone uses area attention at deep stages, a hierarchical bidirectional neck over P1 to P4 fuses context and detail, and the head predicts oriented boxes with distribution focal learning for side offsets and a rotated IoU loss. Training applies a small vertical re bin and a mild reflectance jitter in channel space to resist memorization. We use an interquartile range (IQR) filter to remove noisy and outlier LiDAR points during 3D reconstruction. On KITTI dataset, TriBand-BEV attains 58.7/52.6/47.2 pedestrian BEV AP(%) for easy, moderate, and hard at 49 FPS on a single consumer GPU, surpassing Complex-YOLO, with gains of +12.6%, +7.5%, and +3.1%. Qualitative scenes show stable detection under occlusion. The pipeline is compact and ready for real time robotic deployment. Our source code is publicly available on GitHub.

• The paper introduces a height-aware BEV encoding that segregates LiDAR returns into three distinct height bands, enhancing pedestrian detection.
• It employs a bidirectional high-resolution feature fusion network that effectively combines semantic context and spatial detail for improved small-object localization.
• The approach achieves up to 58.72% pedestrian AP on KITTI at 49 FPS with a lightweight 8.8M parameter model, setting a new benchmark for real-time detection.

TriBand-BEV: Real-Time LiDAR-Only 3D Pedestrian Detection via Height-Aware BEV and High-Resolution Feature Fusion

Introduction and Motivation

TriBand-BEV addresses the challenge of real-time 3D perception for autonomous systems using only LiDAR data, with an emphasis on vulnerable road users such as pedestrians. The architectural and encoding innovations are motivated by the computation and memory limitations inherent to embedded robotics, and by the need for robust detection under occlusion and environmental variability. Prior work in BEV-based LiDAR detection, such as PIXOR and Complex-YOLO, demonstrated the efficiency of 2D projections but either induced high channel complexity or lost critical vertical structure. TriBand-BEV explicitly reformulates 3D LiDAR detection as a 2D BEV problem with rapid post-processing to recover 3D boxes, targeting the operational constraints of real-time deployment.

TriBand-BEV Encoding and Feature Representation

The TriBand-BEV encoding projects the 3D LiDAR point cloud into a compact BEV tensor with three channels, each aggregating maximum reflectance within distinct height bands: $z < 0.65\,\text{m}$, $0.65 \leq z < 1.30\,\text{m}$, and $z \geq 1.30\,\text{m}$, mapped to (R, G, B). By preserving coarse vertical structure, the encoding facilitates object identity and separability, particularly for small objects like pedestrians, whose physical silhouette is distributed across the bands (legs, torso, head). Unlike density or single-height cues, this method is more informative and lightweight, supporting fast inference (tens of kilobytes per frame). Correction for reflectance (amplification and offset) ensures robustness against the sparsity and noise in LiDAR returns.

Network Architecture and Bidirectional Feature Fusion

TriBand-BEV employs a backbone with multi-resolution feature extraction and an area-attention mechanism that yields enhanced contextual aggregation. The backbone produces pyramidal feature maps, and the neck performs bidirectional high-resolution fusion: top-down pathways propagate semantic context to finer resolutions, while bottom-up refinement injects spatial detail for improved localization. The head operates on the fused set of feature maps, favoring improved recall and localization for small BEV footprints, especially pedestrians and cyclists.

Area-attention blocks reduce complexity while preserving non-local spatial dependencies by partitioning feature maps into regions and applying multi-head attention within regions. Lightweight bottlenecks in the neck maintain efficiency alongside the heavier backbone. The detection head predicts oriented bounding boxes using distribution focal learning for side offsets and a rotated IoU loss.

Post-processing and 3D Box Recovery

3D box recovery is accomplished by lifting BEV predictions to LiDAR and camera frames. An interquartile range (IQR) filter removes noisy and outlier LiDAR returns, stabilizing height estimation. Distance-adaptive isotropic dilation and height priors enforce geometric plausibility, particularly in sparse or cluttered environments. Compared to full 3D ground modeling (e.g., RANSAC, GNDNet) which incurs substantial latency, the local IQR filter achieves reliable results at much lower computational cost.

Data Augmentation and Training Strategy

The dataset preprocessing relies exclusively on LiDAR data (KITTI benchmark). For generalization, augmentation procedures apply vertical re-binning and reflectance jittering, exposing the network to local ground variations and preventing memorization of specific environments. The channel-space augmentation preserves reflectance associations, as pixel-wise i.i.d. noise proved detrimental to AP.

The training regime uses SGD with warmup and cosine annealing, distributed data parallelism, and automatic mixed precision. Non-maximum suppression thresholds are tuned for efficient runtime operation.

Numerical Results and Comparative Performance

TriBand-BEV achieves 58.72%, 52.68%, and 47.27% pedestrian BEV AP (easy/moderate/hard) on KITTI, at 49 FPS on a single consumer GPU. These results exceed Complex-YOLO BEV AP by +12.64%, +7.59%, and +3.07% respectively. The model also achieves strong car and cyclist detection, with BEV APs of 81.74% (car/easy) and 57.9% (cyclist/easy). Increasing feature resolution in the neck yields the most substantial accuracy improvements for pedestrians and cyclists, as evidenced by the ablation study.

The entire pipeline operates with 8.8M parameters and maintains real-time inference. The bidirectional high-resolution neck improves recall and localization for small objects (notably pedestrians), with low pedestrian false positives in cluttered urban scenes. Per-class performance degrades for smaller objects at distances exceeding 30m, tied to reduced LiDAR point density, but remains robust for cars owing to denser returns.

Practical and Theoretical Implications

TriBand-BEV underscores the feasibility of lightweight BEV-based detection for embedded robotic systems where compute and power are constrained. The encoding scheme maximizes geometric informativeness while minimizing data dimensionality and latency, making it suitable for onboard deployment. The bidirectional fusion approach demonstrates that high-resolution spatial cues are indispensable for small-object detection, challenging the reliance on coarse features in standard FPNs for BEV inputs.

These innovations provide a template for further research into height-aware BEV encoding and efficient feature fusion for 3D perception in real-world robotic environments. Future developments may include integrating lightweight height refinement networks within object footprints to improve vertical localization without significantly increasing computational overhead.

Theoretically, this work illustrates the impact of encoding design on object separability and detection accuracy, and suggests that compact multi-band representations can outperform conventional density-based or single-height projections in BEV for certain object classes.

Conclusion

TriBand-BEV presents a LiDAR-only detection pipeline that recasts 3D object detection into a 2D BEV formulation augmented with height-aware encoding and bidirectional high-resolution feature fusion. The approach delivers strong numerical gains in pedestrian detection accuracy relative to prior BEV-based methods, with real-time throughput and robust performance under occlusion and clutter. This balance of accuracy and efficiency establishes a new practical baseline for onboard robotic perception. The theoretical underpinnings in encoding and feature aggregation signal opportunities for further advances in small object detection and real-time 3D perception.