DRISHT-E DRone Integrated System for Heterogenic Traffic Extraction

A tool to detect, track and analyze vehicle trajectory from drone footages.

DRISHT-E is a novel offline video based fully-automated traffic data extractor tool for non-lane-based traffic conditions. The objective of this software is to make it compatible with mixed traffic conditions, including

• Signalized Intersections
• Midblock Segments
• Roundabout

Vehicle Detection

Dataset

Drone-based video data was taken from multiple locations at different times of the day with varying geometric and traffic characteristics. The locations include parts of urban cities in India, namely:

• Kerala
• Mumbai
• Gujarat
• Bhopal

A novel dataset of annotated drone footage consists of 186K frames with over 4M annotations. Annotated vehicle categories include 2-Wheeler , 3-Wheeler, Car, LCV and Heavy-Vehicle (Bus & Truck).

Training

YOLOv5 is used to train on this dataset with 2 Nvidia 1080Ti GPUs. Trained models are exported to torchscript format to be used in the software.

Results

Vehicle Tracking

DRISHT-E uses two trackers to track vehicle detections:

1. CSRT (Channel and Spatial Reliability of Discriminative Correlation Filter)
2. KCF (Kernelized Correlation Filter)

These trackers are interchangeably used to maximise the processing speed and reduce noise while maintaining tracking accuracy.

DRISHT-E Software

The software is created using python using frameworks like

• PyTorch,TorchVision: Model inference & processing
• OpenCV: Video & Image operations
• Numpy, Scipy: Numeric operations
• Pandas: Trajectory data analysis
• Streamlit: Visualization

User Interface

Info

This project is currently under development. Much of the data, especially in training, results, algorithm are redacted due to the ongoing research.