Multiple object tracking with data association and correlation filter based on convolutional neural network features

Object tracking allows determination of object positions in successive images without object detection. The kernelized correlation filter performs fast and efficient object tracking in Fourier space. Using deep features that are produced by convolutional neural networks with correlation filter increases tracking success. Three experimental studies were carried out in this thesis. In the first study, performance comparison of different object tracking methods was performed using OTB-100 dataset. In the second study, the effect of different appearance models on tracking success was analyzed for kernelized correlation filter. In the third study, detecting and tracking pedestrians, and associating the tracked and detected pedestrians using 2D MOT 15 dataset were tried. Yolo was used to detect pedestrians. In order to track the detected pedestrians, histogram of oriented gradients (HOG) based kernelized correlation filter was run. Deep features, color histogram, and bounding boxes were used to associate pedestrians. Single object tracking performances of the methods were computed with success, precision, and frames per second metrics, while multi-object tracking performances were computed with metrics such as a combined error rate of detections and number of identity changes. The results showed that the success of object tracking increases when the kernelized correlation filter is used with deep features. Appearance models, which are mostly retrieved from the last layers of convolutional neural networks, enable effective object tracking. Real-time object tracking is possible when the kernelized correlation filter is used with HOG features. Associating pedestrians with deep features in multi-object tracking reduces number of identity switches.