Skip to Main content Skip to Navigation

Vision par ordinateur poursuivi automatique et caracterisation decomportement de civelles

Abstract : Monitoring living organisms makes it possible to understand the dangers that threaten hem and that might have disastrous effects on biodiversity. As part of the MIRA research federation (Milieux and Aquatic Resources) at the University of Pau et Pays de l'Adour, the biology and computer laboratories work together to study migratory fishes in the Adour basin.One of the research actions concerns the acquisition of knowledge about the migratory behavior of an endangered species : the glass eel. One hypothesis to be verified is that the migratory behavior of glass eels is linked to their energy reserves and to the rate at which they use these reserves, because they do not feed during their estuarine migration. For this study, glass eels are marked with a colored elastomer, and then introduced into an experimental medium simulating the tides, and filmed for several weeks in order to follow their movements. The objective of this work is to exploit computer vision techniques allowing automatic tracking of glass eels in an aquarium in order to extract information meeting the needs of biologists, such as counting the number of passages according to the direction of the tide, and measuring the swimming speed and direction. The current techniques for analyzing video data are either entirely manual, or based on elementary image processing that does not explicitly exploit the temporal dimension of the video sequences collected and gives only summary information. The problem of the thesis therefore concerns the contribution of a temporal processing of videos for motion detection, motion estimation and monitoring of glass eels robust to occlusions. Moreover, an unsupervised technique (without heavy training on a large dataset) is desirable, and also obtaining more detailed additional information not accessible to observation by the human operator is envisioned (such as the swimming undulation behavior). In this thesis, the steps of an algorithm allowing to detect and follow glass eels in video sequences are developed. Elver detection uses a background subtraction technique followed by a connected-component analysis of the resulting image in order to extract the geometric information from the bounding boxes of the markings. For color classification, the choice of a hue-luminance-saturation system, combined with an unsupervised classification by the K-means algorithm on the hue component, allows learning of the color range of the classes. These ranges are used to identify the color of each marking detected. The tracking algorithm uses a Kalman filter and a color data association method. Thereafter, a tagging coupling technique allows the individual identification of each glass eel. This makes it possible to obtain their trajectory in the aquarium, as well as dynamic information (average speed, motion direction, number of passages). In parallel, the potential interest of motion estimation by optical flow techniques to extract more precise information is investigated. The velocity-vector field obtained by differential methods (algorithms of Horn & Schunck and of Lucas & Kanade) makes it possible to obtain information currently not available to biologists, such as the swimming undulation of the elvers resulting in the divergence and convergence of the velocity-vector field. Finally, we present the complete information system with the human-machine interface for automatic monitoring of glass eels developed to meet the needs of biologists. This makes it possible to identify glass eels, to determine their direction of passage (with or against the water current) and to count them, thus reducing the tedious observation work time of a human operator.
Document type :
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Thursday, September 22, 2022 - 12:04:38 PM
Last modification on : Friday, September 23, 2022 - 5:15:14 AM


Version validated by the jury (STAR)


  • HAL Id : tel-03783675, version 1



Nawal Eldrogi. Vision par ordinateur poursuivi automatique et caracterisation decomportement de civelles. Bio-informatique [q-bio.QM]. Université de Pau et des Pays de l'Adour, 2021. Français. ⟨NNT : 2021PAUU3063⟩. ⟨tel-03783675⟩



Record views


Files downloads