OPIS - Major subsystem integration and maturation 

Context in source publication

Context 1

... and will deliver significant cost and time savings to the oil industry. The objectives of the initial offering will be to detect and localize large scale damage and anomalies; future offerings will expand on this capability. This effort leverages mature, existing capabilities from across Lockheed Martin which are enhanced, customized and integrated to deliver a new capability. Anomaly, or more generally, change detection is achieved by building an in situ model of the oil platform using point cloud data from the 3D imaging sonar and comparing it with an a priori model of the same platform to detect and localize changes. The accuracy of the in situ constructed 3D model of the platform, and therefore, the accuracy of the change detection, is very sensitive to the knowledge of the sensor pose. AUV’s typically navigate using an inertial navigation system (INS) that uses a variety of navigation aids including GPS, doppler velocity logs (DVL), acoustic positioning systems , depth sensors , and speed of sound sensors. The navigation accuracy of these INSs is typically adequate for most AUV missions. This paper describes the development and implementation of a feature based navigation capability that allows the AUV to navigation safely in the vicinity of an underwater platform while also increasing the accuracy of the in situ constructed 3D model of the platform and change detection. This capability is achieved by fusing the 3D sonar based pose estimate with the INS pose and using this fused pose for the guidance of the vehicle as well as to aid the INS and compensate for drifts. The operational concept for the OPIS is to have an AUV autonomously inspect an offshore drilling platform with minimal user input - the user simply chooses the platform and specifies how much of the platform is to be inspected. The AUV autonomously plans the inspection path around the platform, executes this path to collect sonar data, builds a 3D model of the platform in real-time, and executes change detection to identify anomalies. Feedback in the form of the path of the AUV and the detected anomalies is provided to the operator. The in situ 3D model of the platform constructed from the current inspection along with 3D models of the anomalies are available to the operator upon recovery of the AUV. These models can be exported to a variety of formats to address the needs of individual users. Future version of the OPIS will autonomously plan a re-visit mission to gather close up optical imagery of the sites identified as anomalous. OPIS brings a diverse set of capabilities from across Lockheed Martin together into an integrated system which is capable of achieving the primary goal of autonomously detecting and localizing platform anomalies. OPIS is composed of three primary subsystems: the vehicle, autonomous perception – the transformation of sensor data into information; and autonomous response which is responsible for guiding the vehicle safely through the inspection mission. The vehicle, the TM MARLIN AUV is a mature Lockheed Martin product which has been used on multiple missions. LM autonomous perception technologies have been demonstrated in air using LADAR point clouds which is similar in format though not in quality to the data available from a 3D imaging sonar. LM autonomous response technologies have been applied to a wide variety of domains including land, sea surface, and air. Response and perception technologies are modified and adapted to the undersea environment to achieve the goals of OPIS. The integration of the major autonomy components is achieved via well defined interfaces which allow for the independent development of these key autonomy technologies as well as the rapid insertion of plug and play capabilities. These interfaces follow the ASTM F2541 draft autonomy standard [1] where applicable; this standard defines a messaging interface in terms of message content without any regard to transmission protocols or mechanisms. The OPIS uses TCP/IP to implement the messaging interface between the key autonomy components. Message content, frequency, data types, and validity are all defined and maintained in the only common artifact between the three major OPIS subsystems. Communication between each of the subsystems is achieved through the use of a publish- subscribe message passing scheme. At the initiation of the TCP/IP connection between subsystems, each subscribes to the specific messages which contain information it requires. This approach limits the amount of message traffic across the network and eliminates the need to process messages which don’t contain information pertinent to the subsystem. Initial testing and development of the integrated system is performed in a simulation laboratory which simulates vehicle dynamics, 3D sonar data, and inertial navigation data. This lab testing serves to address many integration issues early in the integration process thereby reducing the need for expensive sea trials. Figure 2 illustrates the integration of the major OPIS subsystems and the maturation of the system through a series of test phases. Major technologies, their adaptation and integration into the OPIS are described next. The perception system controls the imaging sonar sensor and processes the point clouds it produces to refine the vehicle pose relative to the prior model of the oil platform. It also integrates the point clouds in their refined poses to produce a 3D model of the oil rig which is then compared to the a priori model to detect changes. As illustrated in Figure 3, the sonar point clouds are aligned to the prior model using a Random Sample Consensus (RANSAC) [2] approach that bears some similarity to iterative closest point methods [3]. Given the refined poses produced by the sensor- based pose alignment, the in situ 3D model is built by collecting the point cloud data in an octree structure [4], as illustrated in Figure 4. The octree stores the second order moments of the points falling in each octree cell, and also keeps track of the empty space traversed by the ray to each point. Combining these two pieces of information, the algorithm determines whether each octree cell is occupied, empty, or unknown. The octree modeling software used in OPIS not only computes the structure of empty and occupied space, but also provide an extensive set of geometric queries that can be used to efficiently process the model. For example the “find closest point” query is used to match point cloud points to the nearest prior model point during the RANSAC process. To detect changes between the prior model and the newly sensed model, the models are compared on a cell-by-cell basis. Locations in which the sensed model contains occupied space and the prior model contains empty space are marked as positive changes, and locations in which the sensed model contains empty space and the prior model contains occupied space are marked as negative changes. Morphological filtering of the resulting changes and filtering based on position (e.g, to remove ground points as specified in the prior model) help to reduce false alarms. The OPIS system has a mid-grade inertial navigation system onboard which provides ~0.8 nautical mile per hour free-inertial performance. Very accurate vehicle pose information relative to a platform local coordinate frame is available via pose alignment which is an integral part of the in situ model building process described above. Since 3D model building and change detection performance is sensitive to the accuracy of vehicle pose, inertial pose and sensor-based pose may be fused to derive a much improved estimate of vehicle pose which is used for vehicle GNC and as an input to the pose alignment process. The OPIS includes a Kearfott SeaDeVil which is an RLG based gyrocompassing INS with a heading accuracy of 1 milli-radian. The INS is able to employ various navigation aids including a GPS and a Doppler velocity log. The GPS is used every time the vehicle is on the surface including for initial alignment of the INS. The 600 KHz Teledyne RDI DVL provides good estimates of AUV velocity relative to the sea floor if the AUV is less than 90 m from the bottom. If the AUV is beyond this bottom-lock range, the DVL is only able to provide the velocity of the AUV through the water. This may be combined with known currents to generate an estimate of AUV ground velocity, but in general, navigation accuracy is significantly affected. With ground velocity, the positional accuracy of the INS is governed primarily by scale factor (along track) and misalignment (across track) errors and is on the order of 0.05% of distance traveled. While this positional accuracy is adequate for many AUV missions it has a significant effect on the quality of 3D model building and change detection accuracy. The Kalman Filter on board the COTS INS is not available for modification although many of its states and covariances are periodically available via an interface. The INS pose and 3D sensor-based pose are fused as shown in Figure 5 below. The INS pose is available at 25 Hz and the sensor-based pose is available at 1-5 Hz and is delayed by up to 2 seconds. To account for this delay, a history of linear and angular velocities from the INS is maintained and these are used to propagate the sensor-based pose forward to the time of the latest INS pose. The sensor-based pose valid at t k is fused with the latest INS pose using a Bayesian combination. We define a pose state: where North , East , and Down represent the position of the vehicle in a local flat earth Cartesian frame and  ,  ,  are Euler angles. We use x to denote the sensor-based pose and x S I to denote the INS pose. We also define an input vector: where, v N , v E , v D are vehicle velocities in the north, east, and down directions, and   ,   ,   are Euler rates; all of which are available from the INS. To write the state transition equations to propagate the ...