Real-Time Instance Segmentation of Traffic Videos for Embedded Devices.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Embedded Systems

Background:

• Real-time instance segmentation is crucial for intelligent transportation systems.
• Existing methods often struggle with performance on resource-constrained embedded devices.

Purpose of the Study:

• To develop a novel, efficient instance segmentation method for traffic videos on embedded systems.
• To improve accuracy and real-time processing capabilities for traffic analysis.

Main Methods:

• Proposed a novel neural network architecture with a multi-resolution backbone and optimized detection/segmentation branches.
• Introduced a new post-processing technique for mask quality evaluation and a label assignment algorithm for training.
• Conducted an ablation study to balance speed and performance, replacing the backbone with a lightweight design.

Main Results:

• The proposed method achieves real-time performance on embedded devices.
• Demonstrated superior performance compared to the You Only Look At Coefficients (YOLAC) algorithm.
• Achieved 31.57 average precision on the COCO dataset and speeds up to 66.25 FPS on Jetson AGX Xavier.

Conclusions:

• The novel instance segmentation method is effective for real-time traffic video analysis on embedded platforms.
• The architectural modifications and training improvements contribute to both high accuracy and processing speed.
• The method offers a viable solution for intelligent transportation systems requiring efficient visual perception.