Multiple Drosophila tracking with heading direction in crossover and touching scenarios

Online tracking of ants based on deep association metrics: method, dataset and evaluation

“…In contrast, the online MOT methods match detections and trajectories frameby-frame, and heavily rely on object detection results due to a short window of the matching process. The Global Nearest Neighbor algorithm (GNN) is widely used for data association [12,13,14], which calculates the metric of information matching between objects in two consecutive frames. The matching criteria are based on the information, including appearance, position, direction [13,14].…”

“…The Global Nearest Neighbor algorithm (GNN) is widely used for data association [12,13,14], which calculates the metric of information matching between objects in two consecutive frames. The matching criteria are based on the information, including appearance, position, direction [13,14]. In the past, Kalman filter [13,14] and particle filter [15,16] were mostly applied to data association process.…”

“…The matching criteria are based on the information, including appearance, position, direction [13,14]. In the past, Kalman filter [13,14] and particle filter [15,16] were mostly applied to data association process. The standard Kalman filter is used to construct a linear model of constant velocity to predict the object position in the next frame [13,14], thereby narrowing the searching range of detected bounding boxes [17].…”

“…In the past, Kalman filter [13,14] and particle filter [15,16] were mostly applied to data association process. The standard Kalman filter is used to construct a linear model of constant velocity to predict the object position in the next frame [13,14], thereby narrowing the searching range of detected bounding boxes [17]. However, in the situation where the state transition and observation model of an object are non-linear, an extended Kalman filter [18] generates more accurate predictions.…”

“…For Drosophila, researchers cut the wings of Drosophila and let them move in a 2D culture dish. They proposed a tracking framework by integrating the planar geometric characteristics of Drosophila body and motion direction [13,14]. Other methods use motion information; thus, the Kalman filter is used to estimate flight status in 3D space, and the Hungarian algorithm is used to match Drosophila under multi-viewpoints [35,18].…”