Biological Water Quality Monitoring Based On Visual Perception And Fish Movement Feature Integration

With the worsening of the environmental pollution,the extreme shortage of water resources has become a universal problem in the world.Pollution of water sources is becoming increasingly serious,and it leads to a significant increase of toxic and harmful substances in domestic water such as drinking water.The quality of polluted water has posed a great threat to human health and life.Therefore,it is urgent to research online water quality monitoring technology.Biological monitoring technology has gradually become a focus of attention for water quality monitoring researchers due to its real-time,enrichment,and comprehensive characteristics.In biological water quality monitoring,fish are important indicator organisms,and their movement characteristics and physiological characteristics directly reflect changes in the water environment and characterize the current water pollution status.Aiming at the problem of water pollution monitoring,this paper proposes a biological water quality monitoring method combining 3D motion trajectory synthesis and integrated learning by the distinctly different sport characteristics and sport modes of fish with normal water quality and abnormal water quality.The specific research contents are as follows:(1)The fish motion video images were collected using cameras from two perspectives,and the Kuhn-Munkres(KM)algorithm was used to match the feature points of the fish targets.Kalman filtering was used to update the current state,and the optimal tracking position was found as the tracking result.The KCF algorithm compensates for the loss in the tracking process and collisions or occlusions during the motion process,which effectively reduces errors caused by lighting,occlusion,and water surface fluctuations.This algorithm can directly obtain the target motion trajectory,and avoid the re-extraction from the centroid points in the image sequence,thus it greatly improve the efficiency.(2)In order to avoid the one-sidedness of the two-dimensional motion trajectory,the pixel coordinates of the three-dimensional trajectory were synthesized into the pixel coordinates in the experiment to provide a more realistic fish swimming trajectory.During the reconstruction process,two cameras we reused to shoot fish movements at different angles.Thus,if the centroid point is close to or superimposed at one perspective,we can jump to another perspective to eliminate the uncertainty of trajectory tracking.The camera distortion was calibrated according to the DLT,then the corresponding points of the image were determined,and the corresponding points were used as the matching points to calculate the three-dimensional pixel coordinates of the fish's motion trajectory.(3)A representative data set of plus or minus samples was selected,and the number of data sets should be roughly symmetrical,so as to extract the characteristic parameters of fish behavior that can reflect the change of water quality.The pre-processed track image data set was input into the Pointnet model,and the further processed motion feature parameter data set was input into SVM and XGBoost for training at the same time,so as to obtain the base classifier that can recognize different water quality.Finally,the base classifier is combined into a strong classifier through integrated learning to obtain a classification model that can distinguish water quality and realize the purpose of abnormal water quality monitoring.The experimental results show that under the pixel coordinates of the fish's three-dimensional trajectory the fusion of the model using ensemble learning can efficiently and accurately reflect the water quality status.In the real water qualityenvironment simulated by the laboratory,the identification rate of water quality reaches more than 95%.