WEpods: Autonomous Shuttles on Public Roads WEpods partners WEpod - PowerPoint PPT Presentation

WEpods: Autonomous Shuttles on Public Roads

WEpods partners

WEpod route

WEpods functional architecture EasyMile EZ10

WEpods global localization

WEpods functional architecture IBEO localization ADASIS e-Horizon Extended ADASIS v2 e-Horizon with lateral position localization Road users & obstacles

WEpods functional architecture

Sensing - Camera

Sensing - Camera

Sensing - Camera

Sensing - Field of View

Sensing – Radar-Camera combination • Radar detection × Unknown type of object  Location of object  Low false negative rate • Visual (pedestrian) detection × Processing of whole image × Unknown visual scale × High false positive rate × Weather conditions etc.  Recognition of object

Sensing – Radar-Camera combination • Radar-Camera Detection & Classification  Location of object  Projection to camera view  Recognition of object • System Architecture • Setup on DrivePX • Radar to Camera projection and visual cropping Deep-learned Convolutional Neural Network • • Network architecture Network training • • Tracking and fusion

Sensing – Setup • Setup on DrivePX: Radar inputs over Aurix CAN interface • Camera inputs over CSI interface • Cropping based on radar to camera projection • NVidia CUDA and CuDNN • Caffe for DNN-library •

Sensing – Radar-Camera combination • Radar – Camera projection Point projection • Object distance • Camera Calibration • Object size • Visual scale •

Sensing – Classification Network architecture: Conv Conv Conv Conv Max Max 9x9 7x7 3x6 3x8 2x2 2x2 Image 128 filters 512 filters 1024 filters 128 filters Class crop 40x100 56x116 Feature Class learning learning

Sensing – Classification Network training: • Robustness to small variations Translation • Scale • Flip • Contrastive loss learning • • Robustness to appearance variations • Contrastive loss learning Boosting • • Continuous learning • Tracking feedback • False classifications retrained

Sensing – Tracking Tracking: • Continuous localization Fusing sources • Increasing robustness • • Short term prediction

Conclussion • Combining radar and camera Deep-learned Convolutional Neural Network • Less false positives • 3D localization • Multiple types of road users • • Future work Combining Visual lane detection and • localization with e-Horizon Pedestrian intent recognition • Road user intent recognition •