Explore software-defined imaging algorithms for energy-efficient object tracking in this tinyML Talks webcast. Delve into adaptive video subsampling techniques that learn to jointly track objects and subsample future image frames in real-time. Examine the evolution of CMOS image sensors towards computational capabilities, including region-of-interest readout and high dynamic range functionality. Discover how these advancements parallel the progress in radio technology, leading to increasingly programmable and configurable image sensors for application-specific needs. Gain insights into experimental results, FPGA-accelerated algorithms achieving video rate performance, and the potential impact on various imaging applications. Learn about the visual SLAM pipeline, optical edge filtering, and reinforcement learning approaches in this comprehensive exploration of cutting-edge imaging technology.
Towards Software-Defined Imaging - Adaptive Video Subsampling for Energy-Efficient Object Tracking
Overview
Syllabus
Introduction
How have people been trying to tackle this problem
Whats one example
Why use cloud computing
Softwaredefined imaging
Softwaredefined radio
Reconfiguring imaging pipeline
The sequence of image processing operations
The image processing stack
Experiments
Lognormal distribution
Logarithmic quantization
Summary
Visual Slam Pipeline
Optical Edge Filtering
AP Vision
Sensor Changes
Object Tracking
ROI
Solution
Software
Algorithm Acceleration
Reinforcement Learning
Network Architecture
Preliminary Results
Conclusions
Questions
Dynamic Talk Sponsors
Taught by
tinyML
