Empowering Pedestrian Safety: The Obstacle Detection App and AI Model Optimization
Yoonjae Yang
AI Application Developer, Nota AI
Introduction
In our rapidly evolving technological landscape, artificial intelligence (AI) has become a transformative force, revolutionizing various industries. One promising application is Edge AI, which enables AI computations to be processed on edge devices, fostering real-time applications without heavy reliance on cloud infrastructure. This article delves into how AI model optimization techniques have made remarkable strides in improving the efficiency of edge AI models. Specifically, we explore how these advancements have enabled the development of our latest innovation—the obstacle detection app—an Android-based solution designed to enhance pedestrian safety.
The Obstacle Detection App: A Lifesaving Innovation
Our obstacle detection app harnesses the power of AI object detection to analyze real-time camera input from users' devices and identify potential obstacles on the road. This unique capability provides audio or haptic feedback, alerting users to potential dangers, especially for pedestrians who have difficulty or are unable to perceive these obstacles visually.
The Challenge of Edge AI Efficiency
Deploying AI models on edge devices, like smartphones, presents unique challenges due to their limited computational capabilities. AI models, particularly those designed for high-performance servers, often demand significant resources, making them unsuitable for low-spec devices. Ensuring seamless performance and efficiency on such devices is critical for practical and widespread adoption.This development addresses the concern of prolonged reliance on smartphones during pedestrian activities, successfully mitigating battery drainage concerns through AI model optimization, which tackles high power consumption requirements.
AI Model Optimization: Unleashing the Potential
To address the efficiency challenge in edge AI, developers have explored AI model optimization techniques. By "lightening" AI models, these techniques reduce computational and memory requirements without compromising performance. The application of AI model optimization has unlocked new possibilities, enabling developers to create AI-powered solutions that deliver exceptional user experiences, even on devices with modest hardware capabilities.
The Power of the Obstacle Detection App
To demonstrate the effectiveness of AI model optimization, we evaluated the performance of three popular AI models used for object detection: YOLOv5s, YOLOv5n, and YOLOv5s with 30% compression. The goal was to identify the most suitable model for our obstacle detection app, particularly for deployment on low-spec devices like smartphones.
The following table presents the results of our evaluation:
Video demonstrations
Small model - YOLOv5s
▪️ Latency: latency of 1000ms or higher.
▪️ Power Consumption: utilizes a power rate of 4 W/s or more.
2. Nano model - YOLOv5n
▪️ Latency: a responsive latency range of approximately 670~800ms.
▪️ Power Consumption: with remarkable efficiency, consumes a power rate of 2.1 W/s.
3. Optimized: 30% Compression - YOLOv5s
▪️ Latency: a rapid response time, oscillating between 450~580ms.
▪️ Power Consumption: excels in energy efficiency, consuming a mere 2.6 W/s.
Conclusion
Our obstacle detection app, powered by advanced AI model optimization techniques, stands as a lifeline for pedestrian safety. By harnessing the capabilities of AI object detection and optimizing AI models for efficiency, this innovative app empowers users to navigate their surroundings with greater confidence and security. The success of AI model optimization exemplifies its transformative potential for edge AI, expanding the horizons of accessibility and efficiency. As we continue to witness advancements in AI optimization, more life-changing applications like the obstacle detection app will emerge, fostering a future where AI technology truly enhances the quality of our daily lives.
