Development of an embedded system for road condition monitoring based on computer vision and deep learning techniques
The RoadEye project proposes the development and demonstration of an integrated application (or system) for real-time road condition monitoring, using a camera and an embedded system, which can be integrated in complete ADAS systems that provide a full range of functions. This application will be able to track and detect the condition of the road surface in real-time, within a distance of 5-to-25 meters from the vehicle, based on computer vision and machine/deep learning techniques. The techniques that will be developed within the project will be able to classify the state of the road into some preselected categories such as normal road, slippery road, road with surface anomalies such as potholes, speed bumps etc.
To achieve its objectives, the RoadEye project has set the following goals:
1. Creation of a complete road image dataset from cars driving on different roads and with different, real life conditions
2. Development of a computer vision technology based on artificial intelligence και deep learning techniques, which will be trained on real-life data as well as on existing road datasets.
3. Exploitation of Irida’s skills on implementing real-time deep learning architectures based on commercial collaborations with big companies such as Qualcomm, Cadence and Xilinx, and adaptation of suchlike architectures within the RoadEye project.
4. Pilot implementation on an embedded system using heterogeneous computing techniques, in order to to optimally exploit available computing resources such as CPU, GPU and DSP.
Participation to VISION 2018 event in Stuttgart, demonstrating DeepAPI and Roadeye products.
EE1.1 - Administrative and technical management of the project (M1-M18)
EE1.2 - Preparation of interim and final progress reports (M1-M18)
EE2.1 - Data collection (M1-M12): Collection and organization of data with different road conditions, by searching for existing bases and on-site data collection. In the first case, bases such as Kitti, CoCo, Pascal and others will be evaluated. Data collection campaigns will also be run.
EE2.2 - Data Preprocessing (M4-M12): Ensure that images / videos taken with different cameras and under different conditions are appropriate material for the development of the RoadEye application. The goal is to ensure maximum application scalability in different conditions. For this purpose, pre-processing and data augmentation techniques will be used.
EE3.1 - Selection of Artificial Intelligence Networks (M4-M9): Different CNN-based networks will be examined for their complexity and suitability for embedded systems implementation.
EE3.2 - Training of Artificial Intelligence Networks (M7-M15): Networks will be trained with deep learning methods and using frameworks such as Caffe or Tensorflow. Variations of networks such as VGG that have shown significant scalability will also be used. Throughout the course of training, emphasis will be placed on the subsequent implementation in real time.
EE4.1 - Embedded system selection (M10-M12): Embedded platforms and SoCs where pilot RoadEye can be implemented with a) Time / consumption performance criteria and b) Commercial exploitation aiming at widely accepted platforms and low cost. Irida Labs has developed commercial applications on platforms such as Qualcomm Snapdragon and NVIDIA Tegra that will be the main choices.
EE4.2 - Development of embedded software for RoadEye (M10-M18): The chosen platform will use heterogeneous computing techniques to maximize the use of available computing resources. The aim is the accuracy of the system in a ~ 5m radius. to be> 90%, while the system can process real-time HD video - 1080p30.
Check and evaluate RoadEye prototype performance with: - Detection rate for different road conditions and in particular for detecting irregularities (puddles) - Error rate for wrong alerts for RoadEye - Computational complexity and real-time consumption - Error detection rates in different conditions than the algorithm training, using different data from the training system
EE6.1 - Publicity Actions (M6-M36): Publicity actions to disseminate project results.
EE6.2 - Commercial exploitation of results (M6-M36): Participation in international exhibitions in Europe, Asia and USA to promote and promote the project's results.
EE6.3 - B2B meetings (M6-M36): Bilateral promotion of project results (B2B meetings) with companies and organizations mainly abroad to further develop the Irida Labs extroversion.