Summary

The aim of this internship is to characterize and extend the usual perception methods so that they output a probabilistic description of their measurement. Beginning with camera-based computer vision algorithms, the study can then be extended to LiDAR.

Internship Context

Object tracking consists of estimating the positions and trajectories of objects in the robot's environment. This is a popular topic in the field of robotics. Indeed, the perception of the environment can give the robot the ability to interact with it (e.g., grab an object) or to avoid dangers (e.g., avoid pedestrians for autonomous vehicles). An accurate perception of the environment is crucial to ensuring safety for both humans and robots.

Previous works carried out in the team translated the score associated with an object detection into a probability of existence, a first step toward the compatibility of this type of data with fusion and tracking methods. However, these fusion methods require a statistical characterization of the geometric measurement error. Thus, the first goal of this internship is to derive from usual methods a probabilistic description of the detection's localization error, enough to build the likelihood that fusion methods require. The evaluation will be based on the exactness of the estimated probability density function. This will require implementing, adapting, and testing these computer vision methods on both simulated and real data. This work can then be extended to multi-camera and multi-sensor (cameras, LiDAR) data fusion.

Achieving these goals will require a brief literature review to select the most promising methods to characterize.

Environment

The internship will take place in the RAP research team (Robotics, Action and Perception, https://www.laas.fr/en/teams/rap/) at LAAS-CNRS, in Toulouse. The research conducted by the RAP team concerns the design, prototyping, implementation and evaluation of algorithms for: visual perception; sensor-based control and navigation; interactive mobile manipulation; multi-sensor data modeling and fusion. As a robotics team, the addressed issues are: the uncontrolled, dynamic and evolving nature of environments and perceptual conditions; uncertainty management; real-time processing requirement; embedding of algorithms, possibly of their hardware support; reproducibility of results.

LAAS’ research activities fall within the domain of Information Sciences and Technologies and address complex systems (artificial and sometimes natural) generally heterogeneous, and at different scales, to devise theories, methodologies and tools for modeling, designing and controlling them.

Profile

[Expected] The candidate should hold a bachelor's degree in robotics, mathematics, computer science (or related fields), and must pursue a master's degree or an engineering degree in the same fields.

[Expected] Good programming skills (Python and/or C++).

[Expected] Good knowledge of machine learning and computer vision techniques.

[Expected] Interest in the field of robotics or computer vision.

[Appreciated] Experience with machine learning libraries such as PyTorch or Scikit-learn.

[Appreciated] Knowledge of robotics and data fusion would be a plus.

Application

Candidates should provide a resume, a motivation letter and Master grades (both Master 1 and 2). After reviewing these documents, a possible interview will be offered to the candidate so that he or she can present his or her experiences and motivation for the subject, as well as a discussion on the subject. The interview may be held in French or English, at the candidate's choice. Applications will be studied on a rolling basis.

LAAS is in a restricted zone (Zone à régime restrictif, ZRR). The selected candidate will need to be approved by the Ministry. If the candidate is rejected, there is no possibility to pursue the internship.