Low-Power Dynamic Object Detection and Classification With Freely Moving Event Cameras.

Area of Science:

• Computer Vision
• Robotics
• Embedded Systems

Background:

• Event-based cameras offer advantages in dynamic scenes but lag in object recognition accuracy and algorithmic maturity.
• Traditional frame-based systems struggle with dynamic camera motion.

Purpose of the Study:

• To develop an energy-efficient, event-based approach for dynamic object detection and categorization.
• To improve accuracy and algorithmic maturity in event-based object recognition systems.

Main Methods:

• Developed an event-based feature extraction method using activity accumulation and Principal Component Analysis (PCA).
• Proposed a backtracking-free k-d tree mechanism for efficient feature matching and selection.
• Implemented the system on a Field-Programmable Gate Array (FPGA) for high performance-to-resource ratio.

Main Results:

• Achieved superior classification performance on real-world event-based datasets compared to state-of-the-art algorithms.
• Demonstrated real-time FPGA performance for object detection in aerial vehicle flight modes, trained with limited data.
• Highlighted drawbacks of frame-based sensors under dynamic motion and compared favorably to deep learning transfer learning.

Conclusions:

• The proposed event-based system offers a viable, low-power solution for dynamic object detection and categorization.
• The feature extraction and classification framework provides insights for adapting to various low-power applications.
• The FPGA implementation enables high performance for resource-constrained scenarios.