Benchmarking 2D Multi-Object Detection and Tracking Algorithms in Autonomous Vehicle Driving Scenarios

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Safe navigation in autonomous vehicles relies on accurate multi-object detection and tracking (MODT).
• Existing MODT methods lack thorough evaluation in real-world road driving scenarios.
• Estimating the position, orientation, and speed of road users is critical for safety.

Purpose of the Study:

• To benchmark modern multi-object detection and tracking methods for autonomous driving.
• To evaluate the effectiveness of 22 MODT method combinations using the BDD100K dataset.
• To identify limitations and areas for improvement in current MODT algorithms for road scenarios.

Main Methods:

• Utilized image sequences from the BDD100K dataset for onboard vehicle camera perspective.
• Implemented and evaluated 22 distinct combinations of multi-object detection and tracking algorithms.
• Developed an experimental framework with metrics to assess individual module contributions and limitations.

Main Results:

• The combination of ConvNext (object detector) and QDTrack (object tracker) emerged as the top-performing method.
• Analysis revealed substantial limitations in current multi-object tracking methods when applied to road driving images.
• Identified specific challenges including multi-class object differentiation and accurate distance estimation.

Conclusions:

• Current multi-object tracking methods require significant enhancement for safe autonomous driving.
• Evaluation metrics should incorporate autonomous driving specifics like multi-class scenarios and target distance.
• Future research must simulate the impact of MODT errors on driving safety to ensure robust performance.