From Near-Sensor to In-Sensor: A State-of-the-Art Review of Embedded AI Vision Systems.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Embedded Systems

Background:

• AI integration in cyber-physical systems is standard, but cloud processing presents latency and power challenges.
• Traditional cloud-based AI vision struggles with data bottlenecks and high energy demands for real-time applications.

Purpose of the Study:

• To review embedded AI vision systems, focusing on near-sensor and in-sensor processing architectures.
• To analyze critical performance metrics for AI-integrated vision systems, including resolution, frame rate, latency, and power efficiency.
• To compare different embedded processing approaches for AI vision pipelines.

Main Methods:

• Comprehensive analysis of AI-integrated vision system characteristics and performance metrics.
• Examination of near-sensor processing architectures with dedicated hardware accelerators.
• Exploration of in-sensor processing solutions integrating computation directly into the sensor.

Main Results:

• Near-sensor systems offer a balance of flexibility and performance for real-time processing.
• In-sensor processing significantly reduces data movement and power consumption by enabling on-chip computation.
• Trade-offs between flexibility, power consumption, and computational performance were identified for each approach.

Conclusions:

• Embedded AI vision systems, particularly in-sensor processing, are crucial for overcoming cloud limitations in cyber-physical systems.
• Further research into embedded AI vision architectures is needed for next-generation machine vision.
• Optimizing embedded AI vision pipelines requires careful consideration of performance metrics and architectural trade-offs.