Perceptron
| Reference number | |
| Coordinator | Volvo Technology AB - GTT/ATR/EES |
| Funding from Vinnova | SEK 13 739 551 |
| Project duration | June 2017 - November 2019 |
| Status | Completed |
| Venture | Electronics, software and communication - FFI |
| End-of-project report | 2017-01942engelska.pdf (pdf, 3250 kB) |
Purpose and goal
The overall goal of the project was to build DL competence and this has been done by fulfilling the next points: 1. a concept & infrastructure for data-driven evolution of DL applications has been developed. 2. A survey and evaluation of training and inference platforms has been done to help during the hardware selection process. 3. Three deep neural networks for object detection, free space detection and lane detection has been also developed. These networks are real-time and are based on state-of-the-art architectures.
Expected results and effects
As a result of the goal fulfilled during the project, the competence within the partners through the DL development chain has been increased. These chain involved knowledge gain in data logging, network architecture, training and validation academic dissertations. A dataset containing 28000 images with objects, lanes and free space annotations has been created. Also a demonstration truck containing the HW and the SW needed for running DNNs has been produced.
Planned approach and implementation
The project was organized in 8 work packages: WP1 Coordination, WP2 Infrastructure, WP3 Data Collection, WP4 Object Detection, WP5 Free Space Detection WP6 Lane Detection, WP7 Overview and evaluation of training and inference platforms WP8 Demonstrator. The methodology of learn, build and measure has been followed to develop and connect the different packages that compose a DL chain. A report describing the results has been created for each package.