ApproxDet: Content and Contention-Aware Approximate Object Detection for Mobiles

Researchers at Purdue University have developed an advanced video analytic system for content and contention-aware object detection that can be integrated via mobile devices, known as AppoxDet. ApproxDet has applications in solving problems pertaining to machine vision and object identification where compute resources are constrained. ApproxDet offers a multi-branch detection kernel that features a data-driven modeling approach based on real-time performance metrics along with a latency scheduler to allow changes to execution branches once objects are detected. Further, a content-aware feature extractor is integrated to determine the height and width of objects while a contention sensor determines resource levels and availability for user connectivity. Applications include autonomous driving, process or manufacturing automation, agriculture, and more.

Technology Validation: This technology has been validated by benchmarking the system against AdaScale and YOLOv3. ApproxDet offered 52% lower latency and 11.1% higher accuracy.

Advantages

-More accurate than existing solutions

-Reduced latency through scheduling and changeable execution branches

-Ideal for mobile applications

-Adaptive user interface

Applications

-Autonomous driving, vehicle and traffic management

-Process/manufacturing automation

-Agriculture

-Medical

-Augmented Reality

TRL: 5

Intellectual Property:

Keywords: ApproxDet, content-aware object detection, contention-aware object detection, advanced video analytics, machine vision, object identification, mobile object detection, reduced latency object detection, autonomous driving, manufacturing automation, Accuracy, Algorithm, Analytics, Augmented Reality, Civil Engineering, Communications and Computing, Computer Programming, Computer Technology, Data, Data Processing, Data Science, Detection Systems, Latency, Mobile Application, Neural Network, Scalable Video Coding, Smart Cameras, Smart Sensors