Calibration for fish behavior binocular visual observation system based on GA-SVR full vision field model

Abstract: For reducing the calibration error and improving the measurement accuracy of fish behavior observation system, a binocular calibration method based on genetic algorithm-support vector regression (GA-SVR) is proposed to solve the problem of image distortion and light refraction, and realize the indirect calibration of underwater large field of view. In this paper, a standard board with 49 uniformly arranged circular targets points developed by HALCON is chosen for calibration. The diameter of the target is 35.5 mm, the center distance of two targets is 70 mm, and the size of the calibration board is 600 mm×600 mm. Furthermore, in order to reduce the dependence on precision instruments for full-view sampling, a sliding track with bidirectional mobile positioning ability is designed, which is manufactured by the high precision CNC(computerized numerical control) tool with accuracy of 0.05 mm. And along the short axis of the track, 17 pairs of slots are machined at 50 mm intervals to locate the calibration board longitudinal moving distance. A transverse moving bar is erected in the slots, and on which 3 pairs of slots are machined at 400 mm intervals. According to the slot positions, the calibration board is moved along the long axis and the moving rod is moved along the short axis respectively. Sample images are acquired at each slot position, so that the image of the calibration board can cover the whole space of fish tank. The calibration board is used as the basis to collect parallax coordinates and world coordinates, and then the complete sample sets are established in the entire effective vision field of the binocular system. The parallax coordinates of the target points are acquired by Halcon operators, and the relative position information of the calibration board is acquired according to the position of sliding track. The samples of target point used in this study include 2 352 targets from all 49 sample images. The SVR is selected to train the sample set, and three decision function with mathematical expression are established with model parameters calculated by the genetic algorithm. In this article the differential evolution(DE) algorithm, particle swarm optimization(PSO) algorithm and genetic algorithm(GA) are chosen to optimize the SVR parameters, and establish the calibration models respectively. The root mean square error(RMSE), single point error and cross-correlation coefficient are used as evaluation indicators. Based on evaluation results the optimal parameters are selected to establish the position calibration model in X, Y, Z axis respectively. Experimental results show that the genetic algorithm has better optimization effect than the other two algorithms. The mean square errors acquired of X, Y and Z axis are 0.959, 0.893 and 4.381 mm, and the correlation coefficients are 0.999 988, 0.999 998 and 0.998 356, respectively. Compared with traditional calibration methods, the single-point mean square error and distance mean square error of proposed method are 1.861 and 0.706 mm, which are lower than that of calibration method in air (single-point mean square error of 27.75 mm; distance mean square error of 10.188 mm) and underwater calibration methods (single-point mean square error of 8.215 mm; distance mean square error of 2.832 mm). This study could provide reference for quantitative methods of fish behavior.