Some of the most challenging behaviors of autonomous robots are related to navigation tasks. According to a policy of safe navigation, the basic level of behaviors are devoted by strategies that allow the obstacle detection and the collision avoidance. Once these tasks have been conveniently solved, the next level is route planning, i.e., the generation of routes or paths that allow the robot to reach specific places in the environment. Some information about the environment is needed for such kind of planning. This information is managed by the robot control system and concerns the fundamental issue of environment mapping. The robot can store a sort of metric data related to the environment, topological information, including relationships between elements of the environment and features, maybe visual, associated to them, or any combination of both.

The aim of our work is to introduce an efficient method for autonomous robot navigation supported by visual topological maps (Maravall et al., 2013a), in which the coordinates of both the robot and the goal are not needed. More specifically, this novel method is meant to drive the robot toward a target landmark using self-semantic location, while simultaneously avoiding any existing obstacle using exclusively vision capacity. By self-semantic location we refer to the concept used generally in topological models that differs from metric navigation models. Therefore, self-semantic location means that the robot control system is able to determine an approximate, semantic self-location when a landmark already presented in the model is detected and recognized. In summary, self-semantic location refers to the cognitive information associated to a particular place of the environment (e.g., “I am at the restaurant,” so that the robot has a specific cognitive framework at this particular place of the environment).

The proposed algorithm is based on a conventional bug algorithm, although in our version we use only visual information as opposite to the classic versions that employ metric information. As it is well-known the bug algorithms are a family of techniques for obstacle avoidance in robot navigation with metric maps in real-time (Lumelsky and Stepanov, 1987; Lumelsky, 2005). These techniques make the robot head toward the goal and, if an obstacle is encountered, it circumnavigates it and remember how close it gets to the goal. Once the obstacle is avoided, the robot returns to the closest point and continues toward the goal. The main drawback of the conventional metric bug algorithms is that they need the knowledge of the robot localization (the hardest constraint) besides the coordinates of the goal in a common reference framework. To update the robot's position coordinates is necessary the use of external positioning systems.

Our method focuses on the search of the target visual landmark based on the entropy maximization of the images captured by the robot (Fuentes et al., 2014), which is used when the robot is in an unknown localization. Our hypothesis is that there is a direct and positive correlation between the entropy of an image and the probability of this image containing one or several objects: the higher the entropy, the higher its probability of containing several objects inside it; and conversely, the lower the entropy, the higher its probability of containing a single object inside it. Afterwards, when the robot has detected a potential landmark, i.e., it is in a known location, a dual architecture is executed with inspiration on cerebellar system of living beings combined with a brain activity. This dual architecture provides the reactive and anticipatory behaviors for the robot autonomous control.

Although the technique can be applied to any type of autonomous robot, we employ an Unmanned Aerial Vehicle (UAV). The application of our algorithm to these flying robots is justified due to occasionally the UAVs only have available an onboard camera as main input sensor for obstacle detection and landmark and location recognition as well as other specific sensors for flying related matters. This is mainly the case with Micro Aerial Vehicles (MAVs).

The rest of the paper is organized as follows. The next section summarizes some related work with the main findings in this work, namely navigation based on topological maps and obstacle detection and avoidance. After the theoretical foundations of the entropy-based search combined with the bug algorithm, we present the experimental work performed for its validation using an UAV. The paper ends with some suggestions for future research work based on the use of the vehicle onboard cameras for vision-based quality inspection and defects detection in different operating environments.

Our proposal involves the use of visual graphs, in which each node stores images associated to landmarks, and the arcs represent the paths that the UAV must follow to reach the next node. Therefore, these graphs can be used to generate the best path for an UAV to reach a specific destination, as it has been suggested in other works. Practically each traditional method used in ground robots for trajectory planning has been considered for aerial ones (Goerzen et al., 2010). Some of those methods use graph-like models and generally they use algorithms such as improved versions of the classic A^* (MacAllister et al., 2013; Zhan et al., 2014) and Rapidly-exploring Random Tree Star (RRT) (Noreen et al., 2016) or reinforcement learning (RL) (Sharma and Taylor, 2012) for planning. RL is even used by methods that consider the path-planning task in cooperative multi-vehicle systems (Wang and Phillips, 2014), in which coordinated maneuvers are required (Lopez-Guede and Graña, 2015).

On the other hand, the obstacle avoidance task is also addressed here, which is particularly important in the UAV domain because of a collision in flight surely implies a danger and the partial or total destruction of the vehicle. Thus, the Collision Avoidance System (CAS) (Albaker and Rahim, 2009; Pham et al., 2015) is a fundamental part of control systems. Its goal is to allow UAVs to operate safely within the non-segregated civil and military airspace on a routinely basis. Basically, the CAS must detect and predict traffic conflicts in order to perform an avoidance maneuver to avoid a possible collision. Specific approaches are usually defined for outdoor or indoor vehicles. Predefined collision avoidance based on sets of rules and protocols are mainly used outdoors (Bilimoria et al., 1996) although classic methods such as artificial potential fields are also employed (Gudmundsson, 2016). These and other conventional well-known techniques in wheeled and legged robots are also considered for being used in UAVs (Bhavesh, 2015).

Since most of the current UAVs have monocular onboard cameras as main source of information several computer vision techniques are used. A combination of the Canny edge detector and the Hough transform is used to identify corridors and staircases for trajectory planning (Bills et al., 2011). Also, feature points detectors such as SURF (Aguilar et al., 2017) and SIFT (Al-Kaff et al., 2017) are used to analyze the images and to determine free collision trajectories. However, the most usual technique is optic flow (Zufferey and Floreano, 2006; Zufferey et al., 2006; Beyeler et al., 2007; Green and Oh, 2008; Sagar and Visser, 2014; Bhavesh, 2015; Simpson and Sabo, 2016). Sometimes optic flow is combined with artificial neural networks (Oh et al., 2004) or other computer vision techniques (Soundararaj et al., 2009). Some of those techniques (de Croon, 2012) are based on the analysis of the image textures for estimating the possible number of objects that are captured. Finally, as in other scopes of research, deep learning techniques are also been used to explore alternatives to the traditional approaches (Yang et al., 2017).

In words the visual bug algorithm can be summarized as follows: “search the target landmark and once it is visible then move always toward the target landmark and circumnavigate any existing obstacle when necessary.”

Besides the implementation of the visual search procedure (which is obviously less critical when the robot controller is provided with the orientation ⊖ to the target landmark), the critical element is the robot odometry (Lumelsky and Stepanov, 1986) and more specifically the robot's dead reckoning or path integration, for this reason the robot planner is then based in the control of the orientation ⊖ toward the target landmark using topological maps.

Figure 1 shows the pseudocode of the visual bug algorithm (Maravall et al., 2015a):

Besides the basic procedure visual search, this algorithm consists of three additional complex vision-based robot behaviors:

B1: “go forward to the goal”

There are defined three basic environment's states or situations:

S1: “the goal is visible”

The visual bug algorithm sets out also hard computer vision problems for both the perception of the environment's states S₁, S₂, and S₃ and for the implementation of the three basic robot's behaviors B₁, B₂, and B₃, so that it is devoted the bulk of the remaining sections of the paper to describe our proposals aimed at solving these specific computer vision problems.

As it was already pointed out in the pseudocode of the visual bug algorithm, and as it is shown in Figure 2, the visual topological map is a connected graph where each node stores images corresponding to the next nodes/landmarks to which it is connected, and toward which the robot must navigate. Besides, the arcs connecting two nodes stores information about the relative orientation, Θ, of one node/landmark with respect to the other. Hence, this hybrid visual and “odometric information” (just orientations and landmarks) can be used by the robot's navigation module for its self-semantic location, allowing the search and detection of the sequential nodes/landmarks belonging to a specific route.

Notice to the existence of unknown states (high error ε) in which the robot uses entropic vision; known states (low error ε) are used when the robot is able to determinate its position with respect to the L_n landmark.

In the sequel a novel method for this self-semantic location task is proposed (i.e., the task of landmark search and recognition; Maravall et al., 2013b) based on the combination of image entropy for landmark search (Search Mode), and a dual feedforward/feedback vision-based control loop (Homing Mode) for the final landmark homing. Figure 3 shows the finite state automaton that models the robot's controller between the Search Mode [S] and the Homing Mode [H] depending on the magnitude of the error ε or the difference images (respect to the landmarks defined in the topological map) and according to the simple heuristic rule: If {error is big} Then {S} Else {H}. As previously mentioned, once the image containing the candidates for the target landmark has been obtained by the entropy maximization process in the Search Mode, the robot's controller switches to the Homing Mode to guide the UAV toward the target landmark (Maravall et al., 2015a).

A hybrid algorithm for the self-semantic location and autonomous navigation of a robot based on entropic vision and the visual bug algorithm has been presented and tested in a scenario corresponding to a hypothetical emergency situation. The proposed algorithm uses a visual topological map to autonomously navigate in the environment. The nodes in the topological map determine a leave-point or a landmark for the self-location of the robot, in which the robot musts re-oriented its navigation in order to reach the goal. Unlike the classic bug algorithms, our algorithm does not require any knowledge about the robot's coordinates in the environment since the robot uses its own self-location method during the navigation to know its position in each iteration.

Based on the experimental results, it is concluded that this hybrid algorithm is highly robust when the robot is around an unknown location. The robustness is provided by the concept of entropic vision and the search of zones with high entropy. It is empirically confirmed the direct and positive correlation between the entropy of an image and the probability of the image of containing several objects inside. The performance when the robot encounters an obstacle during its navigation is acceptable, using the maximization of visual entropy as strategy. In addition, both techniques (the visual bug algorithm and the visual topological maps) together are able to increase the overall solution performance, reducing the number of iterations along the time for reach the goal landmark defined previously.

Future work is planned toward implementation of this hybrid algorithm on other situations in the real world, which an engineering process as self-semantic location of robots is needed: security in building, surveillance of frontiers or critical infrastructure control. We plan to develop further research work concerning the use of the UAV's onboard cameras for vision-based quality inspection and defects detection, taking profit of our experience in vision-based quality inspection and defects detection in the manufacturing industry, where we have introduced the novel concept of the histogram of connected elements as a generalization of the conventional gray level images histogram.

The Authors DM, JdL, and JPF have been working about the paper titled: “Navigation and self-semantic location of drones in indoor environments by combining the visual bug algorithm and entropy-based vision” in the following tasks: definition of abstract, introduction, the dual FF/FB architecture, entropic-vision, and experimental works with drones.