The Feature Learning Based Video Pedestrian Detection

In this thesis our results show that our contributions of this part by improving the"Training set alignment quality of bounding box scores and perfect multi-frame detection",both by manually sanitized the Caltech training annotations and via algorithmic means for the remaining training samples.We study convolutional neural networks for pedestrian detection and describe which parts affect their performance.To report about background/foreground discrimination,we examine convolutional neural network for pedestrian detection and described which aspects affect their performance.Our detailed study,we report top performance on the Caltech dataset,and provide a new sanitized set of training and test annotations.Pedestrian detection is an acknowledged sub-problem that remains a widespread topic in research due to its various applications.We examine the gap between existing state-of-the-art approaches and the"perfect multi-frame detector and alignment of bounding box scores" Stimulated by the new development in pedestrian detection.We examined by creating a human baseline for pedestrian detection(over the Caltech dataset).Our effects described both localization and background-versus foreground errors.To report about localization errors we examine the influence of training annotation noise on the detector performance and show that we can advance even with a small portion of sanitized training data.Although the widespread research on pedestrian detection,Current papers still show significant developments,proposing that saturated point has not yet been reached.We examine the gap between the state of the art and a newly created human baseline.The results show that there is still a tenfold development to be made before reaching human performance.We aim to examine which issues will help to close this gap it shows that how much improvement we can do it in our thesis.