Intelligent Underwater Stereo Camera Design for Fish Metric Estimation Using Reliable Object Matching

Abstract

Precise fish metric estimation is essential in providing intelligent aquaculture farm decisions. Stereo vision has been widely used for size estimation. Still, many factors affect fish metrics accuracy using a low-cost underwater stereo camera, such as distance, ambient lighting, water velocity, and turbidity. Although such a system is affordable and energy-efficient, they are less accurate in estimating depths than its active counterparts. Since power source is always a problem in offshore aquaculture sites, energy-efficient devices are important. To deal with the accuracy problems of the camera, we propose an effective deep-learning-based object matching to optimize the fish metric estimation. In terms of the challenges of the underwater environment, an analysis of the accuracy of the fish 3D position calculation in the aquaculture cage based on the captured stereo camera images is performed. The analysis assumes a known geometrical configuration of the rectified camera system. The critical factor limiting the 3D fish metric estimation accuracy is the resolution of the computed depth maps of fish. An object-based matching is proposed for underwater fish tracking and depth computing to address this issue using reliable convolutional neural networks (CNNs). For each stereo video frame, an object classification and instance segmentation CNN separates the fish objects from their background. The fish objects are then cropped and matched using sub-pixel disparity computation of the video interpolation CNN. The calculated fish disparities and depth values are used for fish metric estimations. We also tracked each fish and computed the metrics across frames. The median metrics are calculated as the final result to reduce the noises introduced by the different gestures of the fish. Furthermore, underwater stereo video datasets with the actual metrics of sampled fish measured by humans are also constructed to verify the effectiveness of our approach. Our proposed method has less than a 5% error rate for fish length estimation.