LiteReconfig: Cost and Content Aware Reconfiguration of Video Object Detection Systems for Mobile GPUs

  • Technology Readiness Level: 6

Abstract

Researchers at Purdue University have developed an adaptive video object detection system (LiteReconfig) intended for systems where there are limited computing resources and low latency requirements. These challenges are particularly relevant on devices using mobile hardware to analyze video streams. LiteReconfig uses a cost-benefit analyzer to schedule switching to new image features as video conditions change, allowing for a more optimal balance of accuracy and computation time. This is done in a content-aware manner, resulting in selections that are tailored to the specific content of the video stream. This system is intended to run on mobile devices and has shown significantly better performance when compared to existing adaptive object detection systems, while running on an NVIDIA AGX Xavier at speeds of up to 50 fps. This technology has applications in the autonomous driving, machine vision, and facial recognition sectors as it allows for improved object detection while minimizing the computational requirements and latency.

Technology Validation: This technology has been validated using the ImageNet VID 2015 benchmark. The model achieved 1.8% to 3.5% greater object detection accuracy and 20.3X to 74.9X lower latency when compared to existing state of the art systems.

Advantages

  • Low latency capability (50 fps)
  • Mobile GPU compatible

Potential Applications

  • Mobile computing
  • Autonomous vehicles
  • Real time object/facial detection

Contact Information

Name: Matthew R Halladay

Email: MRHalladay@prf.org

Phone: 765-588-3469