Description:
- Autonomous driving
- Surveillance
- Object detection; Object tracking; Anomaly detection
Abstract
USC researchers have developed an asynchronous processing-in-pixel-in-memory (P2M) paradigm that improves energy efficiency and decreases processing requirements and data transfer differences without sacrificing significant accuracy. The P2M paradigm consumes almost 2 times less power compared to the current state-of-the-art while maintaining 88.36% test accuracy. This method would allow for massively parallel spatiotemporal analog convolution, enabling event- and difference-based image processing for a sparse detection apparatus to reduce memory and bandwidth requirements.
Benefit
- Improves energy efficiency without substantially sacrificing accuracy
- Robust to physical non-linearities
- Allows for event- and difference-based image detection and processing
Market Application
In the field of edge devices with computer vision, researchers are investigating approaches such as near-sensor, in-sensor, and in-pixel processing to handle large amounts of sensory data with limited computing resources. In-pixel processing, which incorporates computation capabilities within the pixel array, offers high energy efficiency by generating low-level features instead of raw data. While various in-pixel processing techniques have been demonstrated on traditional frame-based CMOS imagers, the potential for applying this approach to neuromorphic vision sensors remains unexplored.
Publications
Kaiser, Md Abdullah-Al, et al. "Neuromorphic-p2m: processing-in-pixel-in-memory paradigm for neuromorphic image sensors." Frontiers in Neuroinformatics 17 (2023): 1144301. https://doi.org/10.3389/fninf.2023.1144301
Stage of Development
- Tested on state-of-the-art neuromorphic vision sensor datasets
- Available for exclusive or non-exclusive licensing