Improving Object Detection Efficiency Using Neuromorphic Visible Light Communication

This abstract has open access
Abstract Summary/Description
Here’s a concise summary/description of the abstract: This work focuses on improving object detection and tracking efficiency using neuromorphic computing and Visible Light Communication (VLC) with the Metavision SDK. The study leverages the SDK's Python-based inference pipeline, which uses a pre-trained TorchScript model to detect and track vehicles and pedestrians. The system processes event-based data, outputs bounding boxes, and provides confidence levels for each detected object. Dual visualization panes display detection and tracking results, showcasing the system's potential for low-power, high-speed performance. The setup involves using event-based cameras or pre-recorded RAW/DAT files, along with extensive pipeline functionalities such as geometric preprocessing, noise filtering, and data association. Parameters like inference thresholds, RoI filtering, and Non-Max Suppression (NMS-IoU) can be adjusted to optimize detection and tracking performance. Performance is influenced by factors such as lighting conditions, camera placement, and lens focus. While the Python implementation is efficient, a faster C++ alternative is also available. This integration of VLC and neuromorphic processing highlights the feasibility of real-time, energy-efficient object detection, with potential applications in smart cities, IoT systems, and autonomous technologies. The study achieves results comparable to traditional methods while reducing energy consumption and improving efficiency.
Abstract ID :
NKDR74
Department of Computer Science, College of Arts and Science
2 visits