Deep learning-based visual object tracking using edge computing on embedded system

State-of-the-art performance achieved by employing deep neural networks in computer vision tasks has attracted a great attention in visual object tracking in recent years. However, that remarkable performance is associated with higher computational cost, and therefore, higher energy consumption in order to run in real-time. Therefore, using deep neural networks on edge devices with limited resources remains a challenging research problem, particularly when real-time action is crucial. In this thesis, a Siamese-based visual object tracker for edge computing applications is proposed. The aim of this work is to reduce computations, thus boosting the tracker by enabling the network to terminate the inference earlier whenever the result is reliable, in addition to skipping the exhaustive multi-scale search needed for scale estimation. The early exiting behavior is achieved by inserting three additional exit branches into the network at which the staged result is tested on pre-defined criteria to decide when to exit. Moreover, quantization-aware training is applied in order to accelerate the model by using lower precision formats with negligible loss in accuracy. The proposed quantized tracker with an adaptive multi-scale search over two scales runs at 68.5 FPS on CPU achieving 4.6 times faster rate of tracking speed compared to the baseline, associated with an acceptable accuracy loss.