3D object detection (3DOD) is central to real-world vision systems and a critical component in the development of perception capabilities for autonomous vehicles (AVs) and mobile autonomous robots. Real-time 3DOD performed at the frame rates required by AVs presents a very difficult engineering problem, though, because it demands significant computing resources from resource-constrained systems. Constraints in AVs include cost, size, power consumption, performance, and accuracy.
While developing 3DOD capabilities for Recogni’s machine-learning (ML) perception chip—an embedded inference engine designed specifically for vehicles and robots—the engineering team developed a software architecture based on inputs from a high-resolution, high-dynamic-range (HDR), high-frame-rate stereo camera pair that meets the unique constraints and requirements for AV-based systems while delivering excellent accuracy.