Mohammad Javadi

Name Mohammad Javadi
Date of birth July 20, 1994
Originally from Tehra, Iran
Languages Persian, English, Turkish
Email mohammadjv94@gmail.com

Topics of Interest

  • Computer Vision
  • Deep Learning
  • Digital Image Processing
  • Computer Vision Neuro Science
  • Bioinformatics
  • Intelligent Robotics
Mohammad Javadi

B.Sc student in Computer Engineering and Information Technology at Amirkabir University of Technology (Tehran Polytechnic).

Diploma in Mathematics and Physics, Allame Helli(2) High School, Tehran, Iran, 2013.

Download PDF CV



Cognitive Robotics Lab - May 2018 - June 2018

Burgard Exploration in ROS-based Virtual Robots

Amirkabir University of Technology, Tehran, Iran

In this work we implemented the Burgard Exploration as a ROS node. This node subscribes to the published information from the Gazebo environment, scanning laser rangefinder, and the other robots, and then, selects a point on a shared map to explore. The communication between robots is done through the ROSCORE.
The source code is developed in a private repository.


Cognitive Robotics Lab - September 2017 - March 2018

Zoom-RNN: A Novel Method for Person Recognition Using Recurrent Neural Networks

Amirkabir University of Technology, Tehran, Iran

In this research, we introduced a novel method to improve Person Recognition task's accuracy on Person In Photo Albums(PIPA) dataset by integration of Recurrent Neural Networks and feature fusion from the different body parts.
The paper was submitted for ECCV2018.


Cognitive Robotics Lab - October 2017 - April 2018

Vehicle Registration Plate Recognition using CNNs, my BSc Thesis

Amirkabir University of Technology, Tehran, Iran

In this project, that was my bachelor thesis, a two-stage system is implemented for detection and recognition of the Persian plates in urban traffic cameras.
For Plate Detection phase, that finds the location of the plates in each frame of the test video, the Faster-RCNN method is used that generates plates' bounding boxes.
Next level, Plate Recognition, that predicts ID of the detected plates and actually behaves as an OCR module, includes two sub-levels:
First, image segmentation and some low-level image processing algorithms are used to find character contours in the plate. Then, a deep convolutional neural network(SqueezeNet) with 22 classes is used to calculate the type of each character.
Implementation of the project is available on Github.

fig1

Bio-Inspired System Design Laboratory - Spring 2017

Humanoid Robot Detection Using Deep Learning: A Speed-Accuracy Trade-off

Amirkabir University of Technology, Tehran, Iran

We compare several two-stage detection systems based on various CNN's and highlight their speed-accuracy trade off. The approach performs edge based image segmentation in order to reduce the search space and then a CNN validates the detection in the second stage. Paper were published in RoboCup Symposium 2017. (paper Link) Codes are available on Github.

fig1

Bio-Inspired System Design Laboratory - Started February 2016, Ended May 2017

Humanoid Robot Localization using Particle Filter

Amirkabir University of Technology, Tehran, Iran

As humans model their environment, humanoids also must do it to be aware of world model and make appropriate decisions in a timely manner. For this purpose, based on statistical analysis(Particle Filter and Monte Carlo methods) and using provided perception data(goals, lines, landmarks, and center circle in a soccer field) by humanoid vision module, we can use relative distances to our self in order to find our location in a soccer field. Source code is not released because of the team rules.

fig1

Hirad Design and Construction Co - started February 2016 , ended February 2017

Mobile Application Development for Home Automation

Advanced Technology Centre of Amirkabir University of Technology, Tehran, Iran

A smart home involves the control and automation of lighting, heating, ventilation, air conditioning, security and etc. In this project, I implemented a cross-platform application(can be used in Android, iOS, BlackBerry and Desktop platform) using Qt library which enables users to control a smart house through wifi.
Details of the product is available here.


Mobin Intelligent System - started October 2015 , ended February 2016

A Base Code for License Plate Detection

Mobin Intelligent System , Tehran, Iran

Detection and recognition of vehicle registration plate in urban traffic cameras were the goals of our system. The constraint of this project was the system processor that was a DSP board and forced us not to use high-level c++ libraries(like OpenCV, Qt and etc).
For mentioned purpose and problem, I implemented a low-level image processing framework which contains some basic functions(matrix operations, filters, convolutions, edge-based methods and etc.), as they are implemented in OpenCV library.
In another level of the work, in the task of plate detection, me and my other teammates studied about HOG(Histogram of Oriented Gradients) and ACF(Aggregate Channel Features) features on image, and we believed that in our task, ACF features are more useful.
As is mentioned above, the processor was a DSP board and it wasn't feasible to use ACF detector (implemented in Matlab by Piotr Dollár et al.) directly. So another team was hired to convert the source code to low-level C++ code.


Amirkabir Robotic Center - started September 2014 , ended June 2017

A High Speed Vision System for Humanoid Soccer Player Robots

Amirkabir University of Technology, Tehran, Iran

In this research, we implemented a high-speed cognition module for AUTMan Humanoid Robotic Team. The outputs(ball, goal posts, lines, center-circle, and obstacles) of the system are considered as inputs to localization and behavior modules. Challenges of the project are:
1- Humanoids suffer from low-frequency processors, like our platform which has an Intel Celeron Dual Core 1.10GHz processor.
2- Although robots in the soccer game field are placed in a high color information environment, many objects in the field have similarity in terms of color. This problem makes us use an edge-based color-aided method.
Our Algorithm's average frequency is 25-30 Hz which seems appropriate for a humanoid robot.

fig1

Publications

  • Azar SM, Azami S, Ghadimi M, Javadi M, Nickabadi A. Zoom-RNN: A Novel Method for Person Recognition Using Recurrent Neural Networks, submitted for ECCV 2018(Rejected) (Link).
  • Javadi M, Azar SM, Azami S, Shiry S, Ghidary SS, Baltes J. Humanoid Robot Detection using Deep Learning: A Speed-Accuracy Tradeoff, International Robocup Symposium, July 2017 (Link).
  • Technical Reports

  • Sadeghnejad S, Baltes J, Ramezani S, Javadi M, Karimi M, Valaei A, Santos J, Poon T, Yazadnkhou B, Alihosseini D, Hosseini MS. AUTMan Humanoid TeenSize Team Description Paper. (Link)
  • Sadeghnejad S, Baltes J, Ramezani S, Karimi M, Javadi M, Karimi M, Valaei A, Ahmadi A, Hosseinmemar A, Santos J, Poon T, Shamshirdar F, Heydari M, Azari B, Behjou S. AUT-UofM Humanoid TeenSize Team Description Paper RoboCup 2015 Humanoid TeenSize Robot League. (Link)
  • Javadi M, Valaee A. A High Speed Vision System for Humanoid Soccer Player Robots, August 2017. (Link)
  • Javadi M, Mir A. Introduction to Convolutional Neural Networks, Presentation in Foundations of Data Mining, June 2017. (Link)
  • Mohammad Javadi

    Tehran, Iran

    E-mail:
    mohammadjv94@gmail.com
    mohammad.javadi@aut.ac.ir

    Mobile: +98 936 481 9275

  • Github
  • LinkedIn
  • AUTMan Humanoid Robotic Team