ConcentrateNet: Multi-Scale Object Detection Model for Advanced Driving Assistance System Using Real-Time Distant...

Area of Science:

• Computer Vision
• Deep Learning
• Autonomous Systems

Background:

• Advanced Driver Assistance Systems (ADAS) face challenges in detecting small, distant objects.
• Current methods struggle to maintain detection confidence for objects at varying distances.

Purpose of the Study:

• To propose an efficient deep learning object detection network, ConcentrateNet, for real-time detection of distant objects.
• To improve the detection of small and faraway objects in vehicular perspectives.

Main Methods:

• ConcentrateNet utilizes a two-stage inference process: initial detection with a large receptive field to predict a vanishing point, followed by cropped image processing near the vanishing point.
• The network architecture is designed for multi-scale object detection and integrates with existing state-of-the-art models.
• A specialized Non-Maximum Suppression (NMS) method merges results from the two inference stages.

Main Results:

• ConcentrateNet achieves significant precision and recall improvements compared to traditional high-resolution input models.
• The proposed network uses lower input resolution and has less model complexity.
• The model was successfully deployed on a low-powered embedded system (NVIDIA Jetson AGX Xavier) for real-time performance.

Conclusions:

• ConcentrateNet offers an efficient and effective solution for real-time distant object detection in ADAS applications.
• The architecture's adaptability and performance on embedded systems make it suitable for autonomous machines.
• This approach enhances the capability of detecting critical faraway objects, contributing to safer autonomous driving.