Visual semantic analysis: toward a smarter urban navigation – AVISE

The objectives of this project are to find a solution to the problem of robustness and to the implementation of the extraction of objects of interest for mapping but also for navigation. This project is part of the resolution of a navigation problem of type Localization and simultaneous SLAM mapping. We propose to no longer consider low-level geometric bitters (corners, points... ) but to focus on high-level semantic landmarks directly representing objects such as road signs, cars, etc...

To do this, we use advances in the field of AI, particularly in neural networks and deep learning. The latter has become the state of the art in the field of computer vision, particularly for the recognition of shapes or objects in images. We used these techniques to extract the high-level landmarks present in the data from the vehicle’s sensors and proposed the integration of these complex landmarks into an iterative estimation process.

The results are promising.

We have developed an urban furniture detector (focused on road signs) allowing the precise, reliable and fast extraction of the parameterized contours of their geometric shapes.

Based on this detection, we showed the possibility of obtaining a 3D representation of these objects by triangulation constrained multiple observations.

Finally, we integrated these detections into an iterative estimation process of the Bayesian filtering type and studied robustification methods of these types of approaches on simple modeling cases.

The perspectives of this work are:
- Large-scale and real-data validation of our approach;
- Extension of Bayesian estimation robustification techniques to more realistic models;
- Integration of semantic objects other than panels (e.g. floor markings);

Ces travaux ont conduit aux publications suivantes :

E. Hrustic, R. Ben Abdallah, J. Vilà-Valls, D. Vivet, G. Pagès and E. Chau-mette«Robust Linearly Constrained EKF for Mismatched Nonlinear Systems,International Journal of Robust and Nonlinear Control, 2020.

E. Hrustic, and D. Vivet.«Using Traffic Signs as Landmarks in Object-oriented EKF-SLAM«The 16th International Conference on Control, Automa-tion, Robotics and Vision (ICARCV 2020), December 13 -15, 2020, Shenzhen,China

E. Hrustic, and D. Vivet.«Quadric-based Traffic Sign Landmarks Initializa-tion for Object-oriented EKF-SLAM«2020 ASILOMAR , November 1-5, 2020,Virtual Conference

E. Hrustic, Z. Xu and D. Vivet.«Deep Learning Based Traffic Signs Boun-dary Estimation«2020 IEEE Intelligent Vehicles Symposium (IV’2020), Octo-ber, 2020, Las Vegas, USA