@conference{2020berlin1,
title = {UrTra2D – Urban Traffic 2D Object Detection Dataset},
author = {Borna Jelić and Ratko Grbić and Mario Vranješ and Milan Z. Bjelica},
doi = {10.1109/ICCE-Berlin50680.2020.9352154},
isbn = {978-1-7281-5885-3},
year = {2020},
date = {2020-11-09},
booktitle = {2020 IEEE 10th International Conference on Consumer Electronics (ICCE-Berlin)},
abstract = {With progress being made in the field of artificial intelligence and especially machine learning, tech and vehicle companies acquired a powerful tool and made a large step towards realisation of a fully autonomous vehicle. Along with the exploding development of more and more powerful hardware, deep learning has become one of the most dominant fields of research in the automotive domain, succeeding the classical computer vision methods. However, to be able to apply deep learning methods to solve a problem, large and appropriate datasets are required in developing a solution, as there is never enough data for deep learning. In this paper, Urban Traffic 2D Object Detection (UrTra2D) dataset is presented, which is intended for training 2D detectors of specific objects common for urban traffic scenes. The data was recorded with an affordable camera mounted inside the vehicle. The dataset contains video sequences and labelled frames of the traffic in the city of Osijek in different weather conditions during both day and night. There are 5 770 labelled frames, totalling in 22 764 labelled objects throughout 11 categories. The UrTra2D dataset is freely available to the research community upon request.},
howpublished = {M33},
keywords = {ADAS, automotive, deep learning, ieeexplore},
pubstate = {published},
tppubtype = {conference}
}

@patent{2020a,
title = {The optimization method of deep learning based on semantic video segmentation on GPUs},
author = {Filip Baba and Dušan Kenjić and Milan Z. Bjelica and Ivan Kaštelan},
year = {2020},
date = {2020-02-19},
number = {P-2020/0197},
howpublished = {M87},
note = {Pending},
keywords = {automotive, consumer electronics, deep learning},
pubstate = {published},
tppubtype = {patent}
}

@patent{2019p2,
title = {A system of software components for isolated execution of an artificial intelligence algorithm for vehicle},
author = {Milan Z. Bjelica and Vladimir Marinković and Miodrag Đukić and Ivan Kaštelan},
year = {2019},
date = {2019-10-01},
number = {P-2019/1098},
howpublished = {M87},
note = {Pending},
keywords = {ADAS, automotive, deep learning, software framework},
pubstate = {published},
tppubtype = {patent}
}

@conference{icceberlin2019_2,
title = {Optimizing Deep Learning Based Semantic Video Segmentation on Embedded GPUs},
author = {Filip Baba and Dušan Kenjić and Milan Z. Bjelica and Ivan Kaštelan},
doi = {10.1109/ICCE-Berlin47944.2019.8966156},
isbn = {978-1-7281-2745-3},
year = {2019},
date = {2019-09-08},
booktitle = {Consumer Electronics - Berlin (ICCE-Berlin), 2019 IEEE 9th International Conference on},
publisher = {IEEE},
abstract = {Decision making in many industries today is being improved drastically thanks to artificial intelligence and deep learning. New algorithms address challenges such as genome mapping, medical diagnostics, self-driving cars, autonomous robots and more. Deep learning in embedded systems requires high optimization due to the high computational demand, given that power, heat dissipation, size and price constraints are numerous. In this paper we analyze several acceleration methods which include utilization of GPUs for most complex variants of deep learning, such as semantic video segmentation operating in real time. Specifically, we propose mapping of acceleration routines commonly present within deep learning SDKs to different network layers in semantic segmentation. Finally, we evaluate one implementation utilizing the enumerated techniques for semantic segmentation of front camera in autonomous driving front view.},
howpublished = {M33},
keywords = {ADAS, automotive, deep learning, ieeexplore},
pubstate = {published},
tppubtype = {conference}
}

@conference{2019_iwssip,
title = {Deep Learning vs. Safety - Practical Approach and Platform Design Perspective},
author = {Milan Z. Bjelica},
url = {http://www.milanbjelica.info/wp-content/uploads/2019/06/program-web.pdf
https://www.youtube.com/watch?v=tJrDJsBiDqg},
year = {2019},
date = {2019-06-06},
booktitle = {Proceedings of 2019 International Conference on Systems, Signals and Image Processing (IWSSIP)},
publisher = {EURASIP},
address = {Osijek, Croatia},
abstract = {Deep Learning is a promising field, allowing an increase in artificial intelligence applications across many fields, ranging from data science, medical, weather, and aerospace to automotive. Applications of computer vision-based deep learning are vastly assisted by modern System-on-Chip architectures, which provide the required parallelism, heterogeneity and interfacing. However, the application of deep learning to safety-critical contexts where human lives might be at stake, such as in self-driving cars, still has many pitfalls. Ongoing academic research tackles transparent AI, in which the correctness of AI is attempted to be reached by design; however, the outcome of this research is still far-fetched. In this talk, we will discuss a practical approach when integrating deep learning vision-based solutions into a safety-critical context, which can be achieved today. We outline an approach which introduces a software/hardware platform design which fosters diversity, with the goal of minimizing risk of critical failures which are induced by AI in decision making.},
howpublished = {M32},
note = {Keynote},
keywords = {ADAS, automotive, deep learning, keynote, market research},
pubstate = {published},
tppubtype = {conference}
}