Tracking of the growth of roots based on multi-object tracking techniques

Details

Supervisors

De Vleeschouwer, Christophe ; Draye, Xavier

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil électricien

Abstract

Multi-object tracking (MOT) is based on the detection of several objects in individual frames at different time instants and then its association across time, obtaining the trajectory of each object. MOT in videos is an important problem in computer vision which has wide applications in various video analysis scenarios, such as visual surveillance, sports analysis, robot navigation, biology, ... This thesis considers MOT applied into a biological scenario, more precisely the analysis of plants root systems. A collection of frames is provided by one single line-scan camera that acquire images periodically. Each image depicts the root of a plant at a given time instant. To ease the detection of the root tips in each image, the plants have been sprayed with water, so that water drops accumulate on each root tip, making their localization easier. Our main purpose is to define a MOT system to track the growing of plant roots, defining the path/trajectory of each root until the plant is completely developed. Our contributions, briefly explained below, are all related to the exploitation of the shortest path obtained for every node in the image in a graph-based framework. The implementation of MOT can be decomposed in two separate phases. On one hand, a detection phase is implemented, where our main objective is to detect the extremity of each root. The best approach is to define a graph in which we can implement the Dijkstra algorithm for the definition of the shortest path to every node on the image. The graph describes the connectivity of the pixels of the images provided, where each white pixel is considered a node. The extremities of the roots are localized through the analysis of the shortest paths, where each extremity corresponds to a local maxima in a small neighborhood. The second phase of the MOT, consist in the association of the detections across time to define the shape of each root. A track is obtained for each extremity, measuring how each root has evolved during its growth period. Due to the low number of miss-detections per frame, we can use a tracking algorithm defined over a small time window, because every drop is present in almost all frames. The best algorithm that suits our necessities is the Hungarian algorithm, which performance is based on the definition of a cost-matrix that defines the relation of detections between consecutive frames. For the definition of the cost-matrix we have defined two different scenarios: - Forward cost. Definition of a Gaussian distribution that predicts the position of a drop in a future frame, given its position in the current frame. - Backward cost. The cost is based on the distance between a past detection and the shortest path connecting a current detection to the source. This cost matrix is referred to as a backward cost, because the distance is measured from detections in the current frame to detections in the past ones.