Convolutional Residual Learning For Pedestrian Detection And Tracking

Visual object tracking have been an interesting computer vision problematic with numerous real-world applications.By definition,it is a procedure of identifying,locating,and defining the dynamic configuration of one or many moving,possibly deformable objects or parts of objects in each frame of one or several cameras.They only require a minor set of training samples on the initial frame to create an appearance model.Nevertheless,some other existing discriminative correlation filters learn the filters distinctly from feature extraction,and update these filters by means of a moving average operation with an empirical weight.These DCF trackers barely profit from the end-to-end training.The main difficulty in Visual object tracking is how to utilize the extremely limited training data to develop an appearance model robust to a variety of challenges including scale variation,background clutter,partial occlusions and motion blur.This project therefore attempts to elucidate these issues through this approach over the current highest accurate generative model,tracking-by-detection based MDNet[1] and the correlation-based CREST,followed with a work on the effect of the convolutional network depth on its precision in the large-scale image recognition setting.At this stage the experimental results on both standard benchmarks,OTB-2013,OTB-2015 and VOT2016[2],show improvements in speed,precision,and robustness on both trackers.