Optimal Configuration of Multi-Task Learning for Autonomous Driving.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous driving requires high-accuracy, real-time image recognition from diverse sensors.
• Current hardware constraints limit the feasibility of processing complex vision tasks simultaneously.
• Multi-task learning (MTL) offers a solution for efficient parallel processing of multiple image recognition tasks.

Purpose of the Study:

• To investigate MTL for optimizing image recognition tasks in autonomous driving under hardware constraints.
• To propose a novel multi-task decision and optimization (MDO) algorithm for enhanced processing speed, accuracy, and memory efficiency.
• To analyze inter-task correlation (ITC) for effective multi-task set (MTS) construction.

Main Methods:

• Developed the MDO algorithm with three steps: MTS selection, shared backbone/subnet training, and network compression.
• Incorporated semi-supervised learning (SSL) for additional performance enhancement.
• Evaluated performance based on object detection mAP, lane detection accuracy, and latency reduction.

Main Results:

• The MDO algorithm effectively constructs MTS with high ITC, crucial for integrated accuracy.
• The proposed MDO algorithm combined with SSL achieved significant improvements.
• Demonstrated a 12% increase in object detection mAP, 15% improvement in lane detection accuracy, and 27% latency reduction compared to prior methods.

Conclusions:

• Integrated accuracy performance in MTL is critically dependent on inter-task correlation (ITC).
• The MDO algorithm provides an effective framework for optimizing MTL configurations for autonomous driving.
• The proposed approach significantly outperforms existing multi-task learning techniques, paving the way for more efficient autonomous systems.