iQDeep - Machine learning for Autonomous Heavy Vehicles

Purpose and goal

The iQDeep project aimed to enhance self-driving at Scania with ML-based algorithms for perception and safety. Highlights of the project include: - Developing Data Selection Tool (DST) and Cloud-based pipeline for efficient data annotation - Creating deployment pipeline for neural networks, improved hardware with GPUs, and inference software - Developing AI-based models for segmentation, detection, tracking, etc. - Research on road network data, leading to 4 publications and a Licentiate - Research on uncertainty in neural networks, leading to a PhD thesis and 8 publications

Expected results and effects

The expected effect is to ease the integration of state-of-the-art data-driven Autonomous Driving technology into Scania vehicles. During the course of iQDeep, we could close the loop of collecting data in vehicles, semi-automatically selecting the most informative data and sending them for annotation, use the annotated data for training our models, setting up the software and hardware for running the trained models and then integrate the trained model to the vehicle to gain/improve the performance of the perception module as well as the downstream tasks in our autonomous pipeline.

Planned approach and implementation

The execution of the project was led by Scania as the main partner, in collaboration from two labs, namely, Computer Vision (CVL) and Automatic Control (RT), at Linköping University. For the internal part at Scania, 4.5 head-counts were allocated to development of data pipeline, setting up the vehicle software and hardware platform, and training and integration of Deep Learning-based models to our perception pipeline. CVL and RT, hired 2 PhD students to assist Scania with carrying out research in two important areas in AD, i.e., Classification in Road Networks and Uncertainty in NNs.