Temporal and Fine-Grained Pedestrian Action Recognition on Driving Recorder Database.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Automotive Safety

Background:

• Fine-grained pedestrian action recognition is crucial for advanced driver-assistance systems (ADAS) to estimate pedestrian intentions.
• Distinguishing subtle actions like walking straight versus crossing is challenging.
• Existing methods struggle with motion analysis in dynamic backgrounds and require large datasets.

Purpose of the Study:

• To develop an effective fine-grained pedestrian action recognition model using limited data.
• To address challenges in analyzing subtle pedestrian motion from vehicle-mounted recorders.
• To enhance pre-crush safety by accurately estimating pedestrian intentions.

Main Methods:

• Evaluated various configurations for fine-grained action recognition without large-scale databases.
• Focused on learning effective recognition models from small-scale datasets.
• Collected two distinct datasets to highlight the research problem.

Main Results:

• Achieved 91.01% accuracy on the National Traffic Science and Environment Laboratory (NTSEL) database.
• Attained 53.23% accuracy on the near-miss driving recorder database (NDRDB).
• Demonstrated significant improvements of +8.28% and +6.53% over baseline two-stream fusion convnets.

Conclusions:

• It is possible to achieve high accuracy in fine-grained pedestrian action recognition with small datasets.
• The proposed methods enhance pedestrian intention estimation for ADAS.
• This research contributes to improving road safety through advanced action recognition.