Core Concepts
Efficient neural architecture search for multi-modal object tracking with low latency and high accuracy.
Abstract
Multiple object tracking is crucial in autonomous driving.
Existing works focus on neural network design for high accuracy, leading to complexity and latency issues.
Proposed method uses NAS to find efficient architectures for real-time tracking with multiple sensors.
Multi-modal framework enhances robustness in object tracking.
Contributions include a constrained NAS method and evaluation on the KITTI benchmark.
Two-stage NAS approach balances accuracy and latency trade-off effectively.
Stats
Experiments demonstrate algorithm can run on edge devices within lower latency constraints, reducing computational requirements.
Achieved 89.59% accuracy close to SOTA methods while keeping latency below 80 milliseconds.
Quotes
"Our proposed algorithm can greatly reduce computational requirements for multi-modal object tracking while maintaining lower latency."
"Our contributions include a constrained neural architecture search method that aims to complete the MOT task within a specified time limit."