Dynamic Occlusion-Predictive Neural Network for Robust Roadside Multi-Vehicle Tracking

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Roadside perception systems offer advantages over onboard sensors but struggle with occlusion in complex traffic.
• Occlusion artifacts lead to tracking failures and incorrect object identification (ID switches).

Purpose of the Study:

• To develop a robust roadside multi-target tracking framework that overcomes severe occlusion challenges.
• To enhance tracking precision and maintain ID continuity during dynamic occlusion events.

Main Methods:

• Proposed a Dynamic Occlusion-Predictive Neural Network with a Transformer-based module to forecast occlusion ratios.
• Integrated occlusion predictions as dynamic weights in the loss function for adaptive error penalization.
• Introduced a GNN-based Spatial Reasoning Module to infer motion patterns of occluded targets using scene-level constraints.

Main Results:

• The proposed framework demonstrated superior performance over state-of-the-art methods on benchmark and self-collected datasets.
• Achieved a 5.1% improvement in Multi-Object Tracking Accuracy (MOTA) compared to the best baseline.
• Showcased significant advantages in high-occlusion scenarios, preserving ID continuity and minimizing tracking failures.

Conclusions:

• The Dynamic Occlusion-Predictive Neural Network effectively addresses occlusion artifacts in roadside multi-target tracking.
• The framework ensures temporally continuous tracking and robust ID preservation even during prolonged visual occlusions.
• Validated efficacy for real-world roadside multi-target tracking applications.