SBIR Phase I: A CVML platform for intelligent machines

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project is that it will enable small and mid-size robotics and industrial vehicle manufacturers to rapidly deploy computer vision and machine learning in their products and make their machines competitive in the global market place. The project will create a multi-vendor real-time vision stack for industrial machines that enables multiple types of machines to be developed using a uniform software interface. Combined with an app store model, such interfaces will enable a developer community to create application-specific solutions with ease, add features to machines via software and essentially create a new market. When deployed by equipment manufacturers, it will have societal impact by reducing the number of forklift and other industrial and construction machine related accidents, deaths and property damage. There are secondary benefits such as reducing the amount spent on worker compensation. The scientific impact will be the enhanced understanding of principled approaches to stereoscopic depth estimation combined with machine learning based object detectors to create integrated vision systems that function in real-time on commodity hardware. This Small Business Innovation Research (SBIR) Phase I project addresses the fact that current solutions for autonomous vehicles and machines use expensive LIDARs and RADARs in conjunction with cameras. Such systems are cost-prohibitive and ill-suited for industrial applications that operate in structured warehouse and manufacturing environments rather than highways. Manufacturers can benefit from a cheaper integrated vision-based system where all the necessary algorithms, software and hardware engineering has already been done for them. The intellectual merit of the project lies in achieving the following research goals. 1) Develop an algorithmic approach to stereoscopic depth estimation that combines quick-to-generate classic features (e.g., edges and corners) with machine learning. 2) Combine machine learning based object detectors with stereoscopic depth to create an integrated vision pipeline that functions in real-time on commodity hardware. 3) Devise methods to train the system more easily by relying on depth and motion features. 4) Address critical operational design considerations such as thermal and power management and understand requirements to maintain mechanical and structural integrity through periods of intense use. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.