In this presentation, we analyze different high-level perception architectures for L3 ADAS using lidar and HD radar data. First, we evaluate standard fusion strategies, highlighting the limitations of early- and late-stage fusion approaches. We then propose a robust mid-level fusion framework, based on clustering and feature extraction level, which ensures significant performance 'goodness' by incorporating lidar accuracy together with movement estimation from HDR. Furthermore, we address the difficulties in multimodal sensing related to managing extreme point-cloud density disparities and mitigating sensor-specific artifacts.

• High-level perception
• Scan data fusion
• Lidar
• HD radar