research in low cost stereo cameras for underwater imaging is currently very active ... calibration stage and a Contrast Limited Adaptive Histogram. Equalization ...
Low Cost Stereo System for Imaging and 3D Reconstruction of Underwater Organisms Luca Mazzei, Lorenzo Corgnati and Simone Marini ISMAR Marine Sciences Institute CNR - National Research Council of Italy La Spezia - Italy Email: {luca.mazzei, lorenzo.corgnati, simone.marini}@sp.ismar.cnr.it I.
E XTENDED A BSTRACT
Besides the traditional image and video systems, methodologies and technologies for the acquisition of stereoscopic images are becoming simple and effective instruments for reconstructing the three-dimensional (3D) model of the investigated scene. Differently from the traditional 2D images, 3D models provide an accurate geometry of the framed scene, where distance, size and volume of the observed subjects can be automatically estimated. Moreover 3D models are not affected by the perspective distortions typical of image and video data. The use of stereo imaging techniques in underwater marine research is mostly exploited in obstacle detection for autonomous vehicles management and in seabed reconstruction [2]. Due to the high cost of 3D commercial imaging devices, the huge potentialities of this approach have not yet been widely investigated, in particular for what concerns marine biology. 3D modeling of biological structures and characters of underwater organisms would give a dramatic boost in the understanding of life forms and scenarios not directly reachable by humans. Driven by this ambitious goal, research in low cost stereo cameras for underwater imaging is currently very active [3]. This paper presents a low cost stereo system for autonomous underwater imaging aimed at the comprehensive analysis of underwater organisms, benthos, and zooplankton through the investigation of reconstructed 3D models of the framed targets. In order to overcome the tight constraints given by the high cost of commercial underwater stereoscopic devices and the high probability of losing these devices during deployments, the proposed system was developed as autonomous, low cost and with the smallest possible size, in order to be easily deployed both on fixed and mobile platforms (e.g. submerged stations, oceanographic moorings, floats, drifters, and AUVs). The hardware architecture is the key core of the developed system, as it has been implemented in order to integrate state-of-the-art software reconstruction approaches, requiring adequate computational capacity and technologically complex electronics, with small size structures and low cost components. The system has been designed to integrate low costs consumer parts and custom prototypes. Two Raspberry Pi general purpose mini PCs are used for the system control, and two Raspberry Pi camera modules perform the image
acquisition. Each camera module is controlled via a software module running in the associated PC. The software module of the system is responsible for the image acquisition and 3D model reconstruction. Image acquisition is operated by a specific software framegrabber based on a client/server paradigm. Each camera acquisition is handled by its server that is triggered by the client program. The client multithread program allows for changing camera acquisition parameters (as exposure time, iris aperture, ISO speed, etc.) and shooting schedule. The client/server architecture also handles the camera acquisition synchronization. Assessment tests returned average delays between the two cameras of the order of the nanoseconds. The 3D reconstruction is based on a typical dense stereo vision pipeline. The intrinsics and extrinsics camera calibration was performed in a controlled condition environment with a checkerboard calibration pattern [5]. After the image acquisition, a preprocessing procedure is performed for each acquired stereo pair. Image dedistorsion, compensation for light refraction in water, rectification based on the model computed in the calibration stage and a Contrast Limited Adaptive Histogram Equalization, CLAHE [1], are performed in order to prepare images to the next processing step. The Disparity Space Image DSI is obtained with a dense stereo Semi-Global Block Matching (SGBM) [4]. Finally a point cloud is extracted with a triangulation transformation from DSI to real coordinates, in order to obtain data for the 3D model reconstruction offline stage. The disparity evaluation function obtained by the stereo system is very dense and enables the complete capability of taking accurate metric measurements within the obtained 3D models. Tests to validate the measurement accuracy of the system have been performed using objects of known size and gave back satisfactory results. The returned metric errors are of the order of millimeters regarding objects of size of the order of centimeters. The small dimension of the implemented system permits its deployment in any kind of enclosure. In particular the system has been tested in a cylindrical enclosure able to reach 40 m depth and in a Vitrovex sphere capable to reach 6000 m depth. Both these enclosures are equipped with a pair of white light LEDs providing the necessary illumination of the framed scene. All hardware components are optimized for low energy consumption. A small size battery pack is thus sufficient for guaranteeing long term deployments, fullfilling the require-
Fig. 1.
3D Model Reconstruction of a coral (left) and with Z Axis color code measurements from blue to red (right).
ments of a compact autonomous system. The system has been tested in both controlled environment and real sea condition. The field testing deployment has been conducted on a coastal seabed in the Ligurian Sea. The test performed in the real sea condition highlighted the critical issues to be addressed in order to optimize and tune the system. In particular, shallow water light reflections caused by the surface, the water turbidity and the configuration of the LEDs pair came out as the major criticalities, acting as main causes of the metric measurements errors. Figure 1 shows the 3D reconstruction of a 10 cm width coral (left) with measurements on Z-axis (right) in color code. Due to its flexibility in tuning and installation, the system enables several investigation approaches, even in harsh environments. The capability of being hosted onboard almost any kind of platform makes the system a promising device for fishes and zooplankton stereo imaging in the water column, for monitoring benthos growth, and characterization of seabed, both in shallow and deep-sea. R EFERENCES [1]
S. M. Pizer, E. P. Amburn, J. D. Austin, et al.: Adaptive Histogram Equalization and Its Variations. Computer Vision, Graphics, and Image Processing 39 (1987) 355-368 [2] Johnson-Roberson, M.; Pizarro, O. et al. Generation and visualization of large-scale three-dimensional reconstructions from underwater robotic surveys. Journal of Field Robotics, Wiley Subscription Services, Inc., A Wiley Company, volume 27, number 1, pp. 21-51, 2010, ISSN 1556-4967 [3] F. Oleari, F. Kallasi, D. Lodi Rizzini, J. Aleotti, S. Caselli, Performance Evaluation of a Low-Cost Stereo Vision System for Underwater Object Detection. 19th World Congress of the International Federation of Automatic Control (IFAC), Cape Town (ZA), Aug. 24-29, 2014. [4] Hirschmuller, H. Stereo Processing by Semiglobal Matching and Mutual Information, PAMI(30), No. 2, February 2008, pp. 328-341. [5] Zhang Z. A Flexible New Technique for Camera Calibration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(11):13301334, 2000.
