CURRENT THESIS

Robots have to plan their path to the destination using the environment maps available to them and look for the visual clues to make sure they are at their destination .For example, if they are instructed to go to a room where there is a chair available, its easy to see all the rooms and look for the chair to decide if this is a destination room.However,it is possible that more than one rooms on the floor have all objects common to them making it difficult to decide which one is the destination.

The only objects (or patterns) which can differentiate between them are those which are not classified by state of the art machine learning algorithms (AlexNet for example).We propose to use human made sketches of such objects (or patterns).The robot will be able to match those sketches with in the rooms and reach the destination .

We would also like to implement the reverse case where the robot has to guide the humans to a destination . We propose using machine learning algorithms to find the most distinctive pattern (or object) of the destination, and guide the human tot the destination using this distinctive feature .

Student : Muhammad Usama Sardar
Supervisor : Dr . Latif Anjum

Mapping and localization using laser scanner are available in ROS . We intend to use these tools to build map of the environment and then use it for robot path planning .

The experiments will be done on real time robot and sensors . Path planning will be done using LTA* and the developed solution will be equipped with obstacle avoidance mechanism to cater for dynamic obstacles encountered during run - time .

Student : Muhammad Waqas
Supervisor : Dr . Latif Anjum

Adversarial examples have been proposed to deceive deep neural networks into misclassifying objects and scenes . A small amount of carefully calculated noise added to the image can deceive the deep neural network . Autonomous robots today rely heavily on visual odometry for localization and path planning .

Most of the visual odometry algorithms are based on handcrafted features such as SIFT, SURF, Harris, etc . These visual odometry algorithms provide the basis for visual SLAM algorithms such as ORB_SLAM . The objective of this research is find out if adversarial examples exist for handcrafted features that can deceive current state of the art visual odometry/SLAM algorithms.

Student : Zohaib Ali
Supervisor : Dr . Latif Anjum

Visual SLAM is a method where a robot moves in an environment and keeps building map of the environment and localizing itself within the map . The biggest problem in visual SLAM is the drift in localization that keeps on increasing with time .

To solve this, a method known as loop closure is used where robot corrects its map and all previous positions when it comes to a place it has already visited . This becomes impossible when scene changes . For example, the robot comes back to a place it has already visited but the scene, being dynamic, has changed .

The robot will not be able to close the loop . Our approach will be based on detecting the dynamic content and come up with a loop closure algorithm that does consider the dynamic content of the scene as part of image for loop closure purposes

Student : Saran Khaliq
Supervisor : Dr . Latif Anjum

The robots can not be trained for all the unforeseen circumstances . How to act in a new environment? What is the purpose of the new object? Recently, it has been shown that robots can learn on the go by passively observing the human behaviour . What to do when there is no human present in the scene? The robot is expected to have a big database of observations related to his home .

If the robot has seen the person sitting on a chair, in a classroom, can the robot predict what is the use of the neighbouring chair? If the person enters/exits the room, through the door, can the robot predict the remaining entrances/exits in the same building? The man made scenes have repeatability . We propose to leverage this repeatability to transfer observations from one place to another .

Student : Khadija Azhar
Supervisor : Dr . Wajahat Hussain

There is recent rise in interest for agile antenna arrays, e . g . , Google Loon and Drone based arrays . These agile arrays have low deployment time and are well suited for adhoc networks . However, the additional flexibility of these arrays comes with the additional cost of maintaining a coordination network for these arrays .

Additional positioning sensors are required for estimating the location of the antenna array configuration . In this work, we propose a decentralized scheme with a simple webcam to estimate the position of agile array elements inn real time . Our novel contribution is the use of computer vision methods with antenna theory .

Student : Tuba Tanveer
Supervisor : Dr . Wajahat Hussain

Fruit flies pose a significant threat to the fruit industry . Traditional methods used to guard against these enemies seem inefficient . Techniques like removing rotten fruits from the farm fields, using predators like ants and male sterilization are manual and resource demanding methods .

Luring the flies using traps with honey, pleasant smell or different lights are passive techniques . In this work, we propose the use of nano , low cost, drones against these fruit flies . We aim to develop a lightweight acoustic sensory mechanism for these nano drones .

We aim to demonstrate that audio data, captured from simple microphones, can be used to discriminate between helpful and harmful flies using machine learning . Additionally, the same audio data can be used to localize the flies . The two above mentioned methods can enable these nano drones to track the harmful flies .

Student : Alia Khalid
Supervisor : Dr . Wajahat Hussain

Robot is made to learn affordance of an object in the environment where affordance determines possible usage of an object . YouTube videos database is provided to robot to observe humans actions in those videos . By observation of human interaction in different scenes from YouTube videos, robot can learn how to behave in any new environment without any assistance .

What to do when there is no human present in the scene? Youtube has big data including people performing different actions . If the robot watches a video of the person sitting on the sofa, can the robot predict the function of the sofa in his home . If the robot watches a video of the person withdrawing a book from the library shelf, can the robot learn how to interact with the book in the library? We propose to leverage the Youtube videos to train for unforeseen circumstances without any teacher’s help .

Student : Shanza Nazir
Supervisor : Dr . Wajahat Hussain

The standard car security systems respond only when you interact with the car in certain places . However, there are ways to hack the security system . All these methods require the robber to interact with the car . We want to develop an intelligent car security system by using the onboard cameras .

We will use recent deep learning pipeline to detect humans around the car . Additionally, we will use face detection to identify the person . An unauthorized person will be detected before he starts interacting with the parked car .

Student : Yasar Hayat
Supervisor : Dr . Wajahat Hussain

Creating summarized environmental maps of robot's experience is very significant for human usage . A robot can visit multiple locations and record those places in its memory . What if a human wants to navigate using the experience of the robot? It is not possible for human to watch all the recorded videos of multiple hours and/or look at hundreds of photos to navigate .

The constraint to available memory also arises here since robots can have a limited memory and it is not feasible to maintain a large space . Our approach addresses these issues by using a single image from a simple RGB camera and covers open door scenarios, is applicable on modeling a large 3 D environment ; not just a single room and is able to determine difference in similarly looking rooms .

Student : Ahmad Faraz Khan
Supervisor : Dr . Wajahat Hussain

To detect the camera pose in MINOS environment by calibrating the MINOS environment . The task is to re - plot the camera pose of the trajectory followed in the scene .

Student : Usama Muddasar
Supervisor : Dr . Wajahat Hussain

Microbes are important for soil . These are part of healthy ecosystem and Long time use of chemical fertilizers and sprays for pesticides will ruin the soil natural futility due to removal of natural microbes . Our aim to design and train a classifier for various kind of microbes (at the time mainly those which can be seen through light microscope) We are training a classifier using neural network techniques for soil data so that the analysis can be made automated ..

Student : Qurat Ul Ain
Supervisor : Dr . Wajahat Hussain

Tracking in SLAM will fail - possibly due to fast motion, poor image quality (e . g . blur), moving into a featureless area, and so on . In such cases it is necessary to have a mechanism to recover the camera pose with respect to the existing map, as soon as it is possible, and then to resume tracking . This is called relocalization and is a crucial part of any useable SLAM system - otherwise its usefulness would end at the first failure .

At its most basic relocalisation can be described as finding a correspondence between the currentcamera image and the existing map . Once a correspondence is established, giving at least an approximate camera location, it should then be possible to re - start frame - to - frame tracking as usual . This generally means relocalisation will need to search over the entire map for a match .

Student : Ans Qureshi
Supervisor : Dr . Latif Anjum

We are working on path planning of robot, if any removable obstacle comes in path of robot while navigating, it will remove the obstacle and continue to plan the path, otherwise it has to re - plan the entire path .

Student : Suneela Zafar
Supervisor : Dr . Latif Anjum

What if you want some parts of your skype video to be hidden? For example you don't want people to see the dirty clothes in your room . Recently, changing content of video has been proposed, using cartoon drawings . Hiding content using cartoon characters puts attention of the very object one wants to remain obscure . How about removing content from the video? Our goal is to use recent deep learning pipeline to generate personal videos where you can automatically hide the content .

Student : Qasim Zia
Supervisor : Dr . Wajahat Hussain

To get information from a person's eye movement about the concealed information . We work on videos to get the information . Person see a video and during its eye movement get information he try to conceal during watching video ..

Student : Badar Hussain
Supervisor : Dr . Wajahat Hussain

The search for compact representations of images and videos has long been a subject of interest for researchers . One way of compressing an image can be to ignore the color content and use methods of colorization of grayscale images at the decoder . Recent developments in deep learning have allowed colorization of gray scale images with high accuracy . A recent colorization system has been trained on millions of color images and produces exceptional results in certain scenarios .

In this research, we study the feasibility of using such deep learning based colorization of images for image compression . Three different scenarios are proposed and compared for colorization of an image . Different image quality assessment methods i . e . Color Similarity Index Measure (CSIM), Structural Similarity Index measure (SSIM), Peak Signal to Noise Ratio (PSNR) and Mean Square Error (MSE) have been employed to evaluate the impact of the proposed techniques on image compression .

Subjective assessment has been performed to verify the correctness of these metrics as they produce inaccurate assessment in certain scenarios . In video compression, block matching motion estimation is the most computationally expensive and time consuming process . A recent study has presented a method to predict motion from a single image by using Convolutional Neural Networks (CNN) . Using only a single frame, motion of each pixel can be predicted in terms of optical flow .

We analyze whether such a method can be used for accelerating the search process for motion vector calculationWe are working on path planning of robot, if any removable obstacle comes in path of robot while navigating, it will remove the obstacle and continue to plan the path, otherwise it has to re - plan the entire path .

Student : Aroosh fatima
Supervisor : Dr . Wajahat Hussain