ITRI develops autonomous driving vehicle and ADAS technologies, which have the following characteristics:
- Video preprocessing: to enhance quality of perceived image as an adaptation for AI perception capability in various driving conditions. The module includes HDR, image dehazing, and reflection removal.
- Early fusion for heterogeneous sensors: to fuse pixel-wise RGB data with depth information from a plurality of active sensors (i.e. LiDAR/RaDAR). The module includes spatial alignment and depth estimation.
- Performance optimization of AI recognition engine: to design DNN architectures with 7 functions of detection: drive-able area segmentation, vehicle detection, scooter detection, pedestrian detection, traffic sign detection, traffic light detection and lane marking detection that can be embodied into a platform with lower computing budget to achieve up to 30 fps on a single camera framework.
- Late fusion for heterogeneous sensors: to increase perception’s reliability on a multi-sensors platform with a detection filter based on real-time driving environment analysis interpolated against sensor type’s characteristic. The module includes decision fusion.
- Auto generation of testing and training video for corner cases testing: to generate dataset based on computer graphics and/or mixed reality technologies to simulate a customized driving environment based on various parameters, constraints, and scenarios (include NHTSA crash events and Taiwan pre-crash events).
- Full system simulation for autonomous driving vehicle validation: corner case generation, field simulation, CarSim, software in the loop, hardware in the loop, X-by-Wire firmware customization are included to adapt to different vehicle functions, fields, and events. Software and hardware customization provides a highly flexible and reliable simulation.
- AI perception for cross-domain and cross-vehicle: to solve AI model’s bias problem; to develop a methodology to increase model compatibility; to develop a methodology to select critical data samples for AI-based learning; thus reducing resource amount for labeling cost.
Applications & Benefits
- Prompting automotive electronics module and system manufacturers to build cost-effective, cross-domain AI-enabled modules.
- Incorporating ITRI’s AI perception and software-hardware integration capabilities for the development of autonomous vehicles.
- Working with the industry to build an open platform for modular smart EVs.