Sin título

Página 1

Goal: Develop a cost-effective and accurate vehicle speed estimation system using monocular video data.

Why It's Important: This approach addresses the limitations of traditional hardware-based systems like LiDAR and GPS, which are costly and not easily scalable.

Research Problem:

Limitations of Traditional Methods:

High costs and complexity of deployment.

Reliance on expensive hardware like LiDAR, GPS, and radar.

Proposed Solution: A software-based approach using video data and machine learning, providing a scalable alternative.

Methodology 1: YOLOv3 and Optical Flow

Components Used:

YOLOv3: For real-time vehicle detection.

DeepSORT: For tracking detected vehicles across frames.

Pyramidal Lucas-Kanade Optical Flow: To calculate speed using pixel displacement and virtual intrusion lines.

Performance Metrics:

Achieved a Mean Absolute Error (MAE) of 3.38 km/h.

Mean Square Error (MSE) also minimized.

Methodology 2: Deep Homography Transformation Network

Components Used:

Deep Homography Transformation: Converts monocular views to bird's-eye perspectives for better accuracy.

YOLOv4: Enhances vehicle detection capabilities.

CSPDarkNet53: Used for feature extraction to handle occlusion issues.

Performance Improvements:

MSE reduction from 14.02 to 6.56 on datasets like BrnoCompSpeed and AIC18.

Challenges and Solutions:

Occlusion Issues: Addressed using bird's-eye transformation and advanced detection networks.

Complexity of Tracking: Utilized DeepSORT for robust tracking across frames.

Environmental Factors: Adaptability to various lighting and weather conditions through data augmentation and robust model training.