30 credits – Machine Learning for Autonomous Driving

30 credits – Machine Learning for Autonomous Driving

Scania genomgår nu en transformation från att vara en leverantör av lastbilar, bussar och motorer till en leverantör av kompletta och hållbara transportlösningar.

Ingress:
Scania is one of the world’s leading manufacturers of heavy trucks and busses. Over a million Scania vehicles are in active use around the world. At the department of Autonomous Transport Systems (ATS) - Pre-Development & Research, we develop a full stack prototype system for autonomous driving. A thesis project in this department is a great opportunity to work on the forefront of autonomous vehicle development and an excellent way of making contacts for your future working life. Many of our current employees have started their careers with a thesis project. We also have plenty of opportunities for both PhD- and expat positions.

Background:
As a part of the ATS department, the AI team develop AI based functions for our autonomous driving vehicle pipeline. Most of our work currently circle around computer vision but we have plenty of ground to cover also outside of this domain, in the machine learning area. We in the AI-Team also believe in having fun at work so be prepared for lot of fussball, pool, and other team activities with an awesome crew.

Topics to include in the assignment:
Depending on the applicants background, strengths and interest, we are flexible in the exact definition of the thesis. Below are a number of topics we are interested in exploring within the AI team.
  • Motion detection using optical flow. We have an interest in investigating what we can achieve by using optical flow to detect movement between different image frames and thereby identify moving objects. There is a famous neural network called FlowNet which is trained on synthetic (computer graphics) images to estimate optical flow. Deploy this network (or a modified, improved or an alternative one) on our camera set and propose a method for moving object detection.
  • Inference on pre-processed images. Multiple camera feeds from cameras around the vehicle are collected from cameras with different lens opening angles and properties. Investigate, propose and evaluate different pre-processing methods for increased accuracy and generalization and decreased computational load and corresponding neural network adaptions for processing of these image streams. E.g. rectification or image stitching compared to batch processing, etc.
  • Image data compression effects. When capturing images for annotation, or when using publicly available benchmark data sets, they are usually compressed, e.g. as .jpg or .png. But when running inference in a real time automotive system the raw image stream from the cameras are preferably used. The impact on training with compressed images and running inference on raw images, as well as how to tackle this most efficiently, seems to not be very well covered in the literature. To analyse this and propose suitable image compressions is of high interest for us.
  • Network pruning. Propose, implement and evaluate pruning methods for neural networks to decrease the computational load of the target hardware, but with kept accuracy of the predictions.
  • ML uncertainty - data selection tool. The accuracy of a prediction from a neural network (or other ML algorithm), when having no ground truth to compare with, is not well investigated. This is especially true during real-time inference. A sub-application of such a method is to use it for selecting data to annotate in order to expand the training set. The task is to propose, implement and evaluate methods for selecting new training data. The ultimate goal is to feed new images into a tool, with the ML algorithm in the loop, that selects the data with the worst performance from the ML algorithm and use this information to improve the algorithm, e.g. by annotating this data and add it to the data set.
Education:
Master (civilingenjör) in computer science, physics or similar, preferably with specialization in computer vision, machine learning, artificial intelligence, control or robotics.
Number of students: 2 students
Start date: February 2019
Estimated time needed: 20 weeks

Contact persons and supervisors:
Alireza Razavi PhD, supervisor, 08 553 708 39, alireza.razavi@scania.com
Mohammad Nazari PhD, supervisor, 08 553 705 92, mohammad.nazari@scania.com
Mikael Johansson, AI team leader, 08 553 533 45, mikael_v.johansson@scania.com

Application:
Enclose CV, cover letter and transcript of records.


Recruitment will happen continuously and the positions may be filled before the final application deadline.
Mer info
Område Södertälje
Yrkesroll Teknik & Ingenjör, Civilingenjör
Typ av anställning Heltid, Projekt- / Visstidsanställd, Examensarbete
Hemsida http://www.scania.com
Sista ansökningsdag 14 dec (32 dagar kvar)

Om arbetsgivaren

Scania är en världsledande leverantör av transportlösningar. Tillsammans med våra partners och kunder leder vi övergången till ett hållbart transportsystem. Under 2017 levererade vi 82 500 lastbilar, 8 300 bussar samt 8 500 industri- och marinmotorer till våra kunder. Vi omsatte närmare 120 miljarder kronor, varav 20 procent utgjordes av servicerelaterade tjänster. Scania grundades 1891 och finns idag representerat i mer än 100 länder, och har drygt 49 000 medarbetare. Forskning och utveckling är koncentrerad till Sverige, med filialer i Brasilien och Indien. Produktion sker i Europa, Latinamerika och Asien, med regionala produktcentra i Afrika, Asien och Eurasien. Scania ingår i Traton Group. För ytterligare information, besök www.scania.com.