High-Profile VRU Detection on Resource-Constrained Hardware Using YOLOv3/v4 on BDD100K

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Systems

Background:

• Vulnerable Road User (VRU) detection is critical for advanced driver-assistance systems (ADAS) and autonomous vehicles.
• Current computer vision limitations hinder complete VRU detection accuracy and efficiency.
• Processing capacity and bandwidth constraints pose challenges for real-time detection.

Purpose of the Study:

• To enhance VRU detection performance using YOLOv3 and YOLOv4 object detection models.
• To evaluate the impact of novel activation functions (MISH, SWISH) and data augmentation techniques on detection quality.
• To analyze performance-throughput trade-offs for practical implementation.

Main Methods:

• Training YOLOv3 and YOLOv4 models on the BDD100K dataset for VRU and general on-road object detection.
• Retraining models with MISH and SWISH activation functions, replacing Leaky ReLU.
• Implementing data augmentation techniques like mosaic and cutmix, and exploring grid size configurations.

Main Results:

• Trained YOLO models showed significant detection quality improvement over generic MS-COCO models with minimal increase in processing time.
• MISH and SWISH activation functions yielded notable performance gains compared to Leaky ReLU.
• Mosaic, cutmix augmentation, and optimized grid sizes provided cumulative improvements.

Conclusions:

• YOLOv3 and YOLOv4, when trained on specific datasets and optimized, offer superior VRU detection capabilities.
• Advanced activation functions and data augmentation techniques are effective in improving object detection performance.
• The study provides valuable insights into optimizing deep learning models for enhanced road safety applications.