Autonomous vehicle guidance in unknown environments

Gaining from significant advances in their performance granted by technological evolution, Autonomous Vehicles are rapidly increasing the number of fields of possible and effective applications. From operations in hostile, dangerous environments (military use in removing unexploded projectiles, survey of nuclear power and chemical industrial plants following accidents) to repetitive 24h tasks (border surveillance), from power-multipliers helping in production to less exotic commercial application in household activities (cleaning robots as consumer electronics products), the combination of autonomy and motion offers nowadays impressive options. In fact, an autonomous vehicle can be completed by a number of sensors, actuators, devices making it able to exploit a quite large number of tasks. However, in order to successfully attain these results, the vehicle should be capable to navigate its path in different, sometimes unknown environments. This is the goal of this dissertation: to analyze and - mainly - to propose a suitable solution for the guidance of autonomous vehicles. The frame in which this research takes its steps is the activity carried on at the Guidance and Navigation Lab of Sapienza – Università di Roma, hosted at the School of Aerospace Engineering. Indeed, the solution proposed has an intrinsic, while not limiting, bias towards possible space applications, that will become obvious in some of the following content. A second bias dictated by the Guidance and Navigation Lab activities is represented by the choice of a sample platform. In fact, it would be difficult to perform a meaningful study keeping it a very general level, independent on the characteristics of the targeted kind of vehicle: it is easy to see from the rough list of applications cited above that these characteristics are extremely varied. The Lab hosted – even before the beginning of this thesis activity – a simple, home-designed and manufactured model of a small, yet performing enough autonomous vehicle, called RAGNO (standing for Rover for Autonomous Guidance Navigation and Observation): it was an obvious choice to select that rover as the reference platform to identify solutions for guidance, and to use it, cooperating to its improvement, for the test activities which should be considered as mandatory in this kind of thesis work to validate the suggested approaches. The draft of the thesis includes four main chapters, plus introduction, final remarks and future perspectives, and the list of references. The first chapter (“Autonomous Guidance Exploiting Stereoscopic Vision”) investigates in detail the technique which has been deemed as the most interesting for small vehicles. The current availability of low cost, high performance cameras suggests the adoption of the stereoscopic vision as a quite effective technique, also capable to making available to remote crew a view of the scenario quite similar to the one humans would have. Several advanced image analysis techniques have been investigated for the extraction of the features from left- and right-eye images, with SURF and BRISK algorithm being selected as the most promising one. In short, SURF is a blob detector with an associated descriptor of 64 elements, where the generic feature is extracted by applying sequential box filters to the surrounding area. The features are then localized in the point of the image where the determinant of the Hessian matrix H(x,y) is maximum. The descriptor vector is than determined by calculating the Haar wavelet response in a sampling pattern centered in the feature. BRISK is instead a corner detector with an associated binary descriptor of 512 bit. The generic feature is identified as the brightest point in a sampling circular area of N pixels while the descriptor vector is calculated by computing the brightness gradient of each of the N(N-1)/2 pairs of sampling points. Once left and right features have been extracted, their descriptors are compared in order to determine the corresponding pairs. The matching criterion consists in seeking for the two descriptors for which their relative distance (Euclidean norm for SURF, Hamming distance for BRISK) is minimum. The matching process is computationally expensive: to reduce the required time the thesis successfully explored the theory of the epipolar geometry, based on the geometric constraint existing between the left and right projection of the scene point P, and indeed limiting the space to be searched. Overall, the selected techniques require between 200 and 300 ms on a 2.4GHz clock CPU for the feature extraction and matching in a single (left+right) capture, making it a feasible solution for slow motion vehicles. Once matching phase has been finalized, a disparity map can be prepared highlighting the position of the identified objects, and by means of a triangulation (the baseline between the two cameras is known, the size of the targeted object is measured in pixels in both images) the position and distance of the obstacles can be obtained. The second chapter (“A Vehicle Prototype and its Guidance System”) is devoted to the implementation of the stereoscopic vision onboard a small test vehicle, which is the previously cited RAGNO rover. Indeed, a description of the vehicle – the chassis, the propulsion system with four electric motors empowering the wheels, the good roadside performance attainable, the commanding options – either fully autonomous, partly autonomous with remote monitoring, or fully remotely controlled via TCP/IP on mobile networks - is included first, with a focus on different sensors that, depending on the scenario, can integrate the stereoscopic vision system. The intelligence-side of guidance subsystem, exploiting the navigation information provided by the camera, is then detailed. Two guidance techniques have been studied and implemented to identify the optimal trajectory in a field with scattered obstacles: the artificial potential guidance, based on the Lyapunov approach, and the A-star algorithm, looking for the minimum of a cost function built on graphs joining the cells of a mesh over-imposed to the scenario. Performance of the two techniques are assessed for two specific test-cases, and the possibility of unstable behavior of the artificial potential guidance, bouncing among local minima, has been highlighted. Overall, A-star guidance is the suggested solution in terms of time, cost and reliability. Notice that, withstanding the noise affecting information from sensors, an estimation process based on Kalman filtering has been also included in the process to improve the smoothness of the targeted trajectory. The third chapter (“Examples of Possible Missions and Applications”) reports two experimental campaigns adopting RAGNO for the detection of dangerous gases. In the first one, the rover accommodates a specific sensor, and autonomously moves in open fields, avoiding possible obstacles, to exploit measurements at given time intervals. The same configuration for RAGNO is also used in the second campaign: this time, however, the path of the rover is autonomously computed on the basis of the way points communicated by a drone which is flying above the area of measurements and identifies possible targets of interest. The fourth chapter (“Guidance of Fleet of Autonomous Vehicles ”) stresses this successful idea of fleet of vehicles, and numerically investigates by algorithms purposely written in Matlab the performance of a simple swarm of two rovers exploring an unknown scenario, pretending – as an example - to represent a case of planetary surface exploration. The awareness of the surrounding environment is dictated by the characteristics of the sensors accommodated onboard, which have been assumed on the basis of the experience gained with the material of previous chapter. Moreover, the communication issues that would likely affect real world cases are included in the scheme by the possibility to model the comm link, and by running the simulation in a multi-task configuration where the two rovers are assigned to two different computer processes, each of them having a different TCP/IP address with a behavior actually depending on the flow of information received form the other explorer. Even if at a simulation-level only, it is deemed that such a final step collects different aspects investigated during the PhD period, with feasible sensors’ characteristics (obviously focusing on stereoscopic vision), guidance technique, coordination among autonomous agents and possible interesting application cases.