Automatic Vehicle Detection using DETR: A Transformer-Based Approach for Navigating Treacherous Roads

Technical Summary: Automatic Vehicle Detection using DETR

Problem Statement
Automatic Vehicle Detection (AVD) in diverse driving environments faces significant challenges due to varying lighting conditions, complex road types, and a wide variety of vehicle classes. Traditional object detection methods, such as YOLO and Faster R-CNN, often rely on handcrafted features and complex pipelines that struggle to generalize across these varied settings, particularly in "treacherous" or cluttered road scenarios common in developing regions. While Convolutional Neural Networks (CNNs) have been the standard, they often exhibit limitations in capturing global context and handling sparse supervision efficiently. This study addresses the need for a more robust detection framework capable of operating effectively in the complex, real-world conditions found on roads in Bangladesh.

Methodology
The research utilizes the BadODD dataset, a comprehensive collection of 9,825 images (5,896 for training, 1,964 for testing) covering nine districts in Bangladesh. The dataset includes 78,943 annotated objects across 13 distinct vehicle classes, captured under diverse day and night lighting conditions.

The proposed methodology involves a comparative analysis and fine-tuning of two distinct architectures:

1. Baseline (YOLOv8m): The authors initially employed the YOLOv8m architecture, known for its computational efficiency and enhanced backbone. However, preliminary results indicated it did not perform adequately on the specific "treacherous" road conditions of the dataset.
2. Proposed Approach (Co-DETR): The study pivots to a Transformer-based approach using the Detection Transformer (DETR) enhanced with Collaborative Hybrid Assignments Training (Co-DETR).
   • Architecture: The model utilizes a SwinTransformer backbone with specific hyper-parameters (embedding dimensions of 192, depths of [2, 2, 18, 2], and a window size of 12).
   • Training Scheme: Co-DETR employs a collaborative hybrid assignment strategy. It introduces multiple parallel auxiliary heads during training to provide more effective supervision and extract positive coordinates, thereby boosting training efficiency. These auxiliary heads are discarded during inference, ensuring no additional computational cost at deployment.
   • Data Preprocessing: To enhance image quality for the model, the authors applied Histogram Equalization, Contrast Limited Adaptive Histogram Equalization (CLAHE), and Gamma Correction to the BadODD image set.

The models were trained on a high-performance workstation (AMD Ryzen 9 5950X, NVIDIA RTX 3090) using the BadODD dataset. While YOLOv8m required 1.2 hours for 10 epochs, the Co-DETR model required 20 hours for the same duration due to the complexity of its transformer architecture and the collaborative training scheme.

Key Results
The study evaluates performance using Mean Average Precision (mAP) with a confidence threshold of 0.4. The results demonstrate a clear superiority of the Co-DETR approach over the YOLOv8m baseline:

• Epoch 1: Co-DETR achieved an mAP of 0.372, compared to YOLOv8m's 0.236.
• Epoch 4: Co-DETR reached 0.418, while YOLOv8m reached 0.255.
• Epoch 9 (Peak): Co-DETR achieved a peak mAP of 0.438, significantly outperforming YOLOv8m's 0.295.

The data indicates that Co-DETR not only starts with higher accuracy but also possesses a stronger capacity for progressive refinement over training epochs, particularly in the diverse and challenging environments represented in the BadODD dataset.

Significance and Claims
The paper claims to be the first study to experiment with DETR for automatic vehicle detection in complex and varied settings, specifically within the context of Bangladeshi road infrastructure. The primary significance of this work lies in:

• Pioneering Application: Validating the practical application of DETR technology for AVD in diverse, non-standard driving environments where traditional CNN-based methods falter.
• Enhanced Mechanisms: Demonstrating that integrating Collaborative Hybrid Assignments Training (Co-DETR) with DETR effectively addresses issues related to sparse supervision and inefficient feature learning.
• Real-World Relevance: Providing a robust framework for AVD that improves accuracy in challenging scenarios, thereby contributing to the advancement of autonomous navigation technology.

The authors conclude that by leveraging the strengths of Transformers to capture global context and set prediction capabilities, this approach offers a promising direction for improving the safety and efficiency of autonomous vehicles in complex driving conditions. They suggest that future work could focus on optimizing Co-DETR for real-time applications and exploring hybrid approaches to further integrate transformer strengths with other deep learning paradigms.