Visual Traffic Analytics

Context & Background

Intelligent Transportation Systems (ITS) are crucial for smart cities in India. Visual sensor networks (CCTV) capture massive quantities of traffic data, but traditional Automated Number Plate Recognition (ANPR) systems fail under occluded, blurry, or low-light conditions. Vehicle Re-identification (VRID) and counting algorithms are needed to plan city infrastructure and manage traffic congestion.

Problems to be Addressed

Vehicle re-identification across non-overlapping cameras is extremely challenging due to viewpoint variation, illumination changes, and similar make/model structures of common cars. Real-time counting under severe congestion requires algorithms that can operate efficiently at the edge.

Aims and Objectives

1. Develop object detection models for vehicle localization under high density.
2. Formulate deep learning-based vehicle re-identification models (VRID).
3. Design spatial counting algorithms that measure the flow of traffic between nodes.

Methodology

The system uses YOLOv4 for high-speed object detection in video frames. For re-identification, deep CNNs (like ResNet) are trained using ranking losses (triplet loss) to learn key visual features. Late fusion combines global color/texture with local orientation-based keypoints. Testing is conducted using real camera feeds from Indian cities.

Expected Outcomes

Deployment of a real-time traffic monitoring prototype in partner cities, automated vehicle re-identification software, and publications in leading computer vision conferences.