Konstantinos Charalampous

On-line deep learning method for action recognition

In this paper an unsupervised on-line deep learning algorithm for action recognition in video sequences is proposed. Deep learning models capable of deriving spatio-temporal data have been proposed in the past with remarkable results, yet, they are mostly restricted to building features from a short window length. The model presented here, on the other hand, considers the entire sample sequence and extracts the description in a frame-by-frame manner. Each computational node of the proposed paradigm forms clusters and computes point representatives, respectively. Subsequently, a first-order transition matrix stores and continuously updates the successive transitions among the clusters. Both the spatial and temporal information are concurrently treated by the Viterbi Algorithm, which maximizes a criterion based upon (a) the temporal transitions and (b) the similarity of the respective input sequence with the cluster representatives. The derived Viterbi path is the node’s output, whereas the concatenation of nine vicinal such paths constitute the input to the corresponding upper level node. The engagement of ART and the Viterbi Algorithm in a Deep learning architecture, here, for the first time, leads to a substantially different approach for action recognition. Compared with other deep learning methodologies, in most cases, it is shown to outperform them, in terms of classification accuracy.

Robot navigation via spatial and temporal coherent semantic maps

The ability of mobile robots to sense, interpret and map their environments in human terms is decisive for their applicability to everyday activities hereafter. Bearing this view in mind, we present here, for the first time, an integrated framework that aims: (i) to introduce a semantic mapping method and (ii) to use this semantic map, as a means to provide a hierarchical navigation solution. The semantic map is formed in a bottom-up fashion, along the robots course, relying on the conceptual space quantization, the time proximity and the spatial coherence integrated into the labeled sparse topological map. A novel time-evolving augmented navigation graph determines the semantic topology of the explored environment and the connectivity among the recognized places expressed by the inter-place transition probability. The robot navigation part is addressed through an interface that facilitates human robot interaction. High level orders are passed to the robots and unfolded recursively, in a top-down fashion, into local navigation data. The performance of the proposed framework was evaluated on long range real world data and exhibited remarkable results. HighlightsLabeled sparse topological map.Augmented navigation graph.Semantic mapping.Robot navigation.

A Multi-Objective Exploration Strategy for Mobile Robots Under Operational Constraints

Multi-objective robot exploration constitutes one of the most challenging tasks for autonomous robots performing in various operations and different environments. However, the optimal exploration path depends heavily on the objectives and constraints that both these operations and environments introduce. Typical environment constraints include partially known or completely unknown workspaces, limited-bandwidth communications, and sparse or dense clattered spaces. In such environments, the exploration robots must satisfy additional operational constraints, including time-critical goals, kinematic modeling, and resource limitations. Finding the optimal exploration path under these multiple constraints and objectives constitutes a challenging non-convex optimization problem. In our approach, we model the environment constraints in cost functions and utilize the cognitive-based adaptive optimization algorithm to meet time-critical objectives. The exploration path produced is optimal in the sense of globally minimizing the required time as well as maximizing the explored area of a partially unknown workspace. Since obstacles are sensed during operation, initial paths are possible to be blocked leading to a robot entrapment. A supervisor is triggered to signal a blocked passage and subsequently escape from the basin of cost function local minimum. Extensive simulations and comparisons in typical scenarios are presented to show the efficiency of the proposed approach.

Robot navigation in large-scale social maps

Integrated robot framework t for navigation in a habituated environment is proposed.The robots behavior obeys the rules of proxemics theory.Exploits the performed actions of humans to infer about its trajectory. As robots tend to establish their presence in every day human environments the necessity for them to attain socially acceptable behavior is a condition sine qua non. Consequently, robots need to learn and react appropriately, should they be able to share the same space with people and to reconcile their operation to mans activity. This work proposes an integrated robot framework that allows navigation in a human populated environment. This is the first work that employs the performed actions of individuals so as to re-plan and design a collision-free and at the same time a socially acceptable trajectory. Expandability is another feature of the suggested mapping module since it is capable of incorporating an unconstrained number of actions and subsequently responses, according to the needs of the task in hand and the environment in which the robot operates. Moreover, the paper addresses the integration of the proposed mapping module with the rest of the robot framework in order to operate in a seamless fashion. The generic design of this architecture allows the replacement of modules with other similar ones, thus providing adaptability with respect to the environment and so on. The method utilizes off-line constructed 3D metric maps organized in terms of a topological graph. During its perambulation the robot is ample to detect humans while it exploits deep learning strategies to recognize their activities. The memorized actions are seamlessly associated with specific rules -deriving from the proxemics theory- and are organized in an efficient manner to be recalled during robots navigation. Moreover, the paper exhibits the differences of the robot navigation in inhabited and uninhabited environments and demonstrates the alteration of the robots trajectory with respect to the recognized actions and poses of the individuals. The system has been evaluated on a robot able to acquire RGB-D data in domestic environments. The human detection and the action recognition modules exhibited remarkable performance, the human detection one was flawless about its decision while the action recognition one confused actions regarding the number of individuals that participate in them. Last, the robot navigation component was proved capable of extracting safe trajectories in human populated environments.

Efficient Robot Path Planning in the Presence of Dynamically Expanding Obstacles

This paper presents a framework for robot path planning based on the A* search algorithm in the presence of dynamically expanding obstacles. The overall method follows Cellular Automata (CA) based rules, exploiting the discrete nature of CAs for both obstacle and robot state spaces. For the search strategy, the discrete properties of the A* algorithm were utilized, allowing a seamless merging of both CA and A* theories. The proposed algorithm guarantees both a collision free and a cost efficient path to target with optimal computational cost. More particular, it expands the map state space with respect to time using adaptive time intervals in order to predict the necessary future expansion of obstacles for assuring both a safe and a minimum cost path. The proposed method can be considered as being a general framework in the sense that it can be applied to any arbitrary shaped obstacle.

Thorough robot navigation based on SVM local planning

A prerequisite for autonomous robot navigation is the extraction of a path that is both efficient and safe in terms of collision. Towards this end, the paper in hand presents a novel local path planning method, incorporating the support vector machines (SVM) theory. The original SVM based module exploits a 2D map of points which are considered to be obstacles, so as to culminate in a collision free path. A unique attribute of the proposed SVM based local path planning algorithm is that it considers the consecutive positions of the global path trajectory, the embodiment of the robot and clusters the obstacles accordingly. Thus, the derived trajectory is a physically constrained path inasmuch as it considers the maximum margin notion of the SVM theory. Instead of providing a purely theoretical approach for local planning assessed using only artificial data, we integrate our local planner into an autonomous navigation system which is evaluated in real-world scenarios in order to show its efficacy. The latter framework firstly constructs a global 3D metric map of the perceived environment and then it converts it into a 2D map upon which a global path planner unrolls. The global map grows incrementally, by registering the collected point clouds over the robots route towards a goal position. Moreover, the navigation is supported by an obstacle detection strategy based on v-disparity images. The system-and, consequently, the presented local path planner-was evaluated in long range outdoors scenarios, navigating successfully within congestive environments. Autonomous robot path planning.SVM based local planning.Global map formation.

Sparse pose manifolds

The efficient manipulation of randomly placed objects relies on the accurate estimation of their 6 DoF geometrical configuration. In this paper we tackle this issue by following the intuitive idea that different objects, viewed from the same perspective, should share identical poses and, moreover, these should be efficiently projected onto a well-defined and highly distinguishable subspace. This hypothesis is formulated here by the introduction of pose manifolds relying on a bunch-based structure that incorporates unsupervised clustering of the abstracted visual cues and encapsulates appearance and geometrical properties of the objects. The resulting pose manifolds represent the displacements among any of the extracted bunch points and the two foci of an ellipse fitted over the members of the bunch-based structure. We post-process the established pose manifolds via

A tensor-based deep learning framework

l_1