Machine vision empowers computers to understand the world through images. This capability has become a crucial component in a myriad of applications, and relies heavily on efficient image sensing for accurate object tracking and recognition. In recent years, advancements in image sensing technology have significantly enhanced the quality of data captured, leading to remarkable progress in machine vision systems. However, the power efficiency of these image sensors has yet to be optimized at the same pace, as existing image sensor models often require high power consumption due to the demanding processes of image frame capture and read-out. As a result, efficient image sensing has become increasingly necessary in order to unlock the full potential of machine vision capabilities.
Researchers at Arizona State University have developed a reinforcement learning (RL) based object tracking method which improves power efficiency of image sensors by filtering out redundant information while sensing. This is achieved by employing a pre-trained convolutional neural network (CNN) and a long short-term memory (LSTM) network, to predict the region of interest (ROI) and subsampling pattern for the consecutive image frames. Based on the application and the trajectory of the object's motion, the user can either utilize the predicted ROI or coarse subsampling pattern to switch off the pixels for sequential frame capture. This method simultaneously outperformed existing methods in tracking efficiency and reduced power consumption during image sensing, presenting an opportunity for optimizing machine vision capabilities in both energy usage and performance.
Related publication: Energy-Efficient Object Tracking Using Adaptive ROI Subsampling and Deep Reinforcement Learning
Potential Applications:
- Autonomous driving
- Security and surveillance systems
- Augmented reality
- Precise object tracking
- Reconfigurable image cameras
Benefits and Advantages:
- Improved tracking precision
- Increased energy efficiency