Harshit Makwana

Transport analytics / Case study

Turning Movement Count Estimation

I developed a traffic-estimation workflow that combined large-scale data preparation, seasonality analysis and tree-based models to estimate vehicle turning movements.

Harshit Makwana

Concept illustration of vehicle trajectories through a four-way road intersection.
AI-generated concept illustration · not a project screenshot
My role
Data engineering, traffic modelling & visualization
Timeline
2022–2023
Focus
Transport analytics

01 / Context

Connecting road data, changing traffic patterns and retrainable models.

Turning-movement estimates need to account for more than a single traffic snapshot. My work examined changes associated with weekdays, festivals and the COVID-19 period, while addressing anomalies and inconsistencies in data collected from different sources.

02 / Ownership

What I built

  • Gathered millions of data points through web scraping, APIs, downloads and paid data services.
  • Analysed seasonality and anomalies, including weekday patterns, festivals and COVID-19-related changes.
  • Used the GEH formula in the outlier-screening workflow and developed preprocessing scripts in PostgreSQL, Trino and PL/SQL.
  • Mapped and connected road records in Python, trained and compared models, and visualized results for clients.
  • Developed a retraining pipeline so new data and changing traffic trends could be incorporated into the modelling workflow.

03 / Engineering

Technical approach

Prepare connected road-level data

I combined data acquisition with database-side preprocessing and Python mapping. This work brought road records and their associated values together before modelling, while seasonality analysis informed how I interpreted unusual observations.

Compare models and support retraining

I evaluated Random Forest and XGBoost and also experimented with CatBoost and LightGBM. I built the surrounding pipeline to make retraining easier as new observations and traffic trends became available, then presented the resulting data through client-facing visualizations.

04 / System view

Workflow at a glance

  1. Traffic sources

    APIs, scraping and downloaded datasets

  2. Preparation

    GEH screening, SQL processing and road mapping

  3. Model evaluation

    Tree-based model experiments

  4. Delivery

    Client visualizations and retraining pipeline

05 / Delivery

Outcomes

Random Forest — reported accuracy
86%
XGBoost — reported accuracy
94%

The project recorded an accuracy score of 86% for Random Forest and 94% for XGBoost. Alongside those results, I delivered preprocessing, mapping, visualization and retraining workflows.

These are the accuracy figures recorded for the project. The metric definition and validation split are not published here, so they should not be read as a standardized benchmark.

Technologies & methods

  • Python
  • PostgreSQL
  • Trino
  • PL/SQL
  • Random Forest
  • XGBoost
  • CatBoost
  • LightGBM

Scalable Data Fusion Engine

As a backend developer, I built data-fusion logic, FastAPI services, ETL workflows and shared security utilities for a system analysing roadway data.

GIS Conflict Detection and Prediction

I built a geospatial ML pipeline to detect conflicts between roadway geometry and trip coordinates, from spatial data preparation to continuous monitoring.

VehiSense Pro

I developed a vehicle-monitoring system with fine-tuned YOLOv8, tracking and event-detection logic, connected to a FastAPI web application through WebSockets.