Sensor Fusion for Traffic Management Intermediate-Fusion#

A BEVFusion-based intermediate-fusion reference implementation that deploys camera and lidar sensor fusion on Intel GPU. The implementation provides two inference pipelines that process synchronized image and point-cloud streams and output 3D object detections in a KITTI-style format.

This implementation features two pipelines tailored to different model architectures:

  • bevfusion: split-model PointPillars pipeline using four ONNX models.

  • bevfusion_unified: unified SECOND pipeline that runs a single ONNX model with custom OpenVINO sparse operations.

Both pipelines target the Intel GPU runtime under OpenVINO and support FP32 and INT8 precision modes.

Key Features#

Discover the key features that set our implementation apart and see how it meets the intermediate-fusion requirements of your intelligent traffic management solution. For a highly performant and cost-efficient solution, leverage the Intel-powered Certified AI Systems. Whether you are evaluating a BEVFusion-based detection pipeline or validating your hardware platform’s capabilities, this reference implementation serves as the perfect foundation.

  • Intel GPU-accelerated inference using OpenVINO, with iGPU and dGPU support for heterogeneous computing configurations.

  • Docker-first workflow for rapid validation: pull the published image and run the automated smoke test in minutes without a native build.

  • Native host build path for production integration, with granular control over precision, dataset, and device selection.

  • INT8 post-training quantization (PTQ) via NNCF, with per-subgraph precision flags for camera backbone, LiDAR PFE, fuser, and detection head.

  • KITTI-format evaluation tooling and dataset conversion utilities for DAIR-V2X-I and KITTI-360.

  • Support for converting NVIDIA CUDA-V2XFusion checkpoints to OpenVINO IR, enabling reuse of existing trained models on Intel hardware without retraining.

Benefits#

  • Enhanced AI Performance: Achieve superior 3D object detection accuracy with intermediate-fusion that tightly couples camera and lidar features before the detection head, outperforming late-fusion approaches under challenging conditions.

  • Accelerated Time to Market: Speed up validation by using the pre-built Docker image and automated smoke-test scripts, reducing environment setup to a single pull-and-run step.

  • Cost Efficiency: Lower your inference costs with INT8 quantization on Intel GPU, maintaining detection quality while significantly reducing compute and power requirements.

  • Simplified Development: Reduce integration complexity with a unified build system, preset dataset configurations, and reference conversion guides for existing NVIDIA checkpoints.