Human beings do not have well defined shapes neither well defined behaviors. In dense outdoor environments, they are as a consequence hard to detect and algorithms based on a single sensor tend to produce lot of wrong detections. Moreover, many applications require algorithms that work very fast on CPU limited mobile architectures while remaining able to detect, track and classify objects as people with a very high precision. We present an algorithm based on the contribution of a range finder and a vision based algorithm that addresses these three constraints: efficiency, velocity and robustness and that we believe is scalable to a large variety of applications.
Detecting and tracking pedestrians accurately is essential to design efficient and robust collision avoidance systems. But traditional approaches to pedestrian detection and tracking in dense urban environments suffer from tracking failures and wrong classifications. We propose in this paper a system that recursively estimates the true outlines of every tracked target using a set of segments called "Appearance". Both the state and the true contours of each target are recursively estimated and can then be used for accurate classification. We show also that accurate information on target outlines allow for a meticulous occlusions computation and an enhanced data association. The performances of this new approach is assessed through a qualitative and quantitative comparison with a state of the art pedestrian detection algorithm.
Because pedestrians have neither well defined shapes nor well defined behaviors, detecting and tracking them from a moving vehicle remains a difficult task. To serve as an onboard driver assistance system, a perception algorithm also needs to be both fast and robust. We present in this paper a system that reaches a good level of reliability by efficiently combining the data of two sensors -a laser scanner and a camera -while remaining tractable on CPU limited mobile architectures.
We present an evolution of traditional occupancy grid algorithm, based on an extensive probabilistic calculus of the evolution of several variables on a cell neighbourhood. Occupancy, speed and classification are taken into account, the aim being to improve overall perception of an highly changing unstructured environment. Contrary to classical SLAM algorithms, no requisite is made on the amount of rigidity of the scene, and tracking do not rely on geometrical characteristics. We believe that this could have important applications in the automotive field, both from autonomous vehicle and driver assistance, in some areas difficult to address with current algorithms. This article begins with a general presentation of what we aim to do, along with considerations over traditional occupancy grids limits and their reasons. We will then present our proposition, and detail some of its key aspects, namely update rules and performance consequences. A second part will be more practical, and will begin with a brief presentation of the GPU implementation of the algorithm, before turning to sensor models and some results.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.