Search

Recently, we have been witnesses of accidents involving autonomous vehicles and their lack of sufficient information at the right time. One way to tackle this issue is to benefit from the perception of different view points, namely collaborative perception. We propose here a decentralized collaboration, i.e. peer-to-peer, in which the agents are active in their quest for full perception by asking for specific areas in their surroundings on which they would like to know more. Ultimately, we want to optimize a trade-off between the maximization of knowledge about moving objects and the minimization of the total information received from others, to limit communication costs and message processing time. To this end, we chose to use Dempster-Shafer Theory (DST) in order to identify different types of uncertainties. In particular, DST allows us to distinguish what has never been perceived (out of range or occluded area) — which is mainly what collaborative perception tries to reduce — from what is debated among different sources (conflict arising from fusion of sensors or other vehicles perceptions). More generally, DST takes into account the specificity of evidence, meaning that it provides information about the reliability of an agent’s belief, which is crucial for safety. DST also features the advantage of easily dealing with data incest with its Cautious fusion rule, which is a problem inherent to the decentralized approach. However, DST comes with high spatial and computational complexities, especially for dealing with data incest in fusion, which limits its usage to random experiments with few possible outcomes. Thus, we first proposed an efficient exact method to compute the decompositions needed for this Cautious fusion, exploiting what we called focal points. Then, we generalized this method to any Möbius transform in any partially ordered set (including all transformations in DST), we found ways to efficiently compute these focal points and we proposed a generalization of the decomposition required by the Cautious fusion. This generalized decomposition allows one to use this Cautious fusion in more cases, in particular cases where an agent has gathered very specific evidence. This enhances both accuracy and computational stability in consecutive fusions. However, algorithms naively based on our formulas would have a higher worst-case complexity than the complexity of the optimal general algorithms commonly employed in DST — which is already more than exponential. Therefore, we later proposed algorithms with complexities always better than the state of the art, and more general, leveraging properties of distributive lattices. After this work on the fusion process itself, we tackled the issue of redundancy and irrelevance in decentralized collaborative perception. For this, we proposed a way to learn a communication policy that reverses the usual communication paradigm by only requesting from other vehicles what is unknown to the ego-vehicle, instead of filtering on the sender side. We tested three different models to be taken as base for a Deep Reinforcement Learning (DRL) algorithm and compared them to a broadcasting policy and a random policy. More precisely, we slightly modified a state-of-the-art generative model named Temporal Difference VAE (TD-VAE) to make it sequential. We named this variant Sequential TD-VAE (STD-VAE). We also proposed Locally Predictable VAE (LP-VAE), inspired by STD-VAE, designed to enhance its prediction capabilities. We showed that LP-VAE produced better belief states for prediction than STDVAE, both as a standalone model and in the context of DRL. The last model we tested was a simple state-less model (Convolutional VAE). Experiments were conducted in the driving simulator CARLA, with vehicles exchanging parts of semantic grid maps. Policies learned based on LP-VAE featured the best trade-off, as long as future rewards were taken into account., Récemment, nous avons été témoins d'accidents impliquant des véhicules autonomes et leur manque momentané d'information pertinente. Une manière d'adresser ce problème est d'avoir recours à la perception collaborative, c'est-à-dire de bénéficier de la perception d'une même scène sous différents points de vue. Nous proposons ici une collaboration décentralisée, i.e. pair-à-pair, dans laquelle les agents sont actifs dans leur quête pour la perception complète en demandant des zones spécifiques dans leur voisinage sur lesquelles ils voudraient en savoir plus. In fine, nous voulons optimiser un compromis entre la maximisation du savoir à propos des usagers de la route et la minimisation du volume total d'information reçu des autres, dans le but de limiter les coûts en communications et le temps de traitement des messages.Dans cette optique, nous avons choisi d'utiliser la Théorie de Dempster-Shafer (DST) afin d'identifier différents types d'incertitude. En particulier, la DST distingue ce qui n'a jamais été perçu (zone hors de vue ou occultée) – ce qui est principalement ce que la perception collaborative essaie de réduire – de ce qui est débattu parmi des sources différentes (conflit provenant de la fusion de capteurs ou de perceptions d'autres véhicules). Plus généralement, la DST prend en compte la spécificité des observations, c'est-à-dire qu'elle fournit des informations sur la fiabilité des croyances d'un agent, ce qui est crucial pour la sécurité routière. La DST a aussi pour avantage, avec sa règle de fusion Cautious, d'éviter facilement la consanguinité des données, un problème inhérent à l'approche décentralisée. Toutefois, la DST vient avec de fortes complexités en temps et en espace, particulièrement dans le calcul de la fusion Cautious, ce qui limite son usage à des expériences aléatoires comportant peu d'événements atomiques.Ainsi, notre première contribution fut de proposer une méthode exacte et efficace pour le calcul des décompositions nécessaires à cette fusion Cautious, en exploitant ce que nous avons appelé points focaux. Nous avons ensuite généralisé cette méthode à toute transformée de Möbius dans tout ensemble partiellement ordonné (incluant toutes les transformations en DST), nous avons trouvé des moyens de calculer efficacement ces points focaux et nous avons proposé une généralisation de la décomposition requise par la fusion Cautious. Cette décomposition généralisée permet d'employer cette fusion Cautious dans plus de cas, en particulier ceux où un agent a reporté des observations très spécifiques. Nous montrons que ceci améliore à la fois la précision et la stabilité calculatoire de fusions successives.Cependant, des algorithmes basés naïvement sur nos formules auraient une plus haute complexité de pire cas que celle des algorithmes généraux optimaux communément utilisés en DST – qui est déjà plus qu’exponentielle. De fait, nous avons proposé plus tard des algorithmes ayant des complexités toujours meilleures que celles de l’état de l’art, et étant plus généraux, tirant partie des propriétés des treillis distributifs. Après ce travail sur le processus de fusion en lui-même, nous nous sommes attaqués aux problèmes de redondance et de non-pertinence dans la perception collaborative décentralisée. Pour cela, nous avons proposé un moyen d'apprendre une politique de communication qui renverse le paradigme usuel de communication en ne demandant des autres véhicules que ce qui est inconnu de l'ego-véhicule, au lieu de filtrer du côté émetteur. Nous avons testé trois modèles différents pour servir de base à un algorithme d'apprentissage profond par renforcement (DRL) et les avons comparés à une politique de broadcast et à une politique aléatoire. Plus précisément, nous avons légèrement modifié un modèle génératif de l'état-de-l'art nommé Temporal Difference VAE (TD-VAE) pour le rendre séquentiel. Nous avons nommé cette variante Sequential TD-VAE (STD-VAE).

Dempster-Shafer Theory (DST) generalizes Bayesian probability theory, offering useful additional information, but suffers from a high computational burden. A lot of work has been done to reduce the complexity of computations used in information fusion with Dempster's rule. The main approaches exploit either the structure of Boolean lattices or the information contained in belief sources. Each has its merits depending on the situation. In this paper, we propose sequences of graphs for the computation of the zeta and M\"obius transformations that optimally exploit both the structure of distributive semilattices and the information contained in belief sources. We call them the Efficient M\"obius Transformations (EMT). We show that the complexity of the EMT is always inferior to the complexity of algorithms that consider the whole lattice, such as the Fast M\"obius Transform (FMT) for all DST transformations. We then explain how to use them to fuse two belief sources. More generally, our EMTs apply to any function in any finite distributive lattice, focusing on a meet-closed or join-closed subset. This article extends our work published at the international conference on Scalable Uncertainty Management (SUM). It clarifies it, brings some minor corrections and provides implementation details such as data structures and algorithms applied to DST., Comment: Extension of an article published in the proceedings of the international conference on Scalable Uncertainty Management (SUM) in 2019

Dempster-Shafer Theory (DST) generalizes Bayesian probability theory, offering useful additional information, but suffers from a high computational burden. A lot of work has been done to reduce the complexity of computations used in information fusion with Dempster's rule. Yet, few research had been conducted to reduce the complexity of computations for the conjunctive and disjunctive decompositions of evidence, which are at the core of other important methods of information fusion. In this paper, we propose a method designed to exploit the actual evidence (information) contained in these decompositions in order to compute them. It is based on a new notion that we call focal point, derived from the notion of focal set. With it, we are able to reduce these computations up to a linear complexity in the number of focal sets in some cases. In a broader perspective, our formulas have the potential to be tractable when the size of the frame of discernment exceeds a few dozen possible states, contrary to the existing litterature. This article extends (and translates) our work published at the french conference GRETSI in 2019., Extension of an article published in the proceedings of the french conference GRETSI 2019