STUDY ON THE KEY RESEARCH ON THE FUNDAMENTAL TECHNOLOGIES OF PEDESTRIAN DETECTION APPLYING DEEP LEARNING.

Abstract

Numerous forms of intelligent technology rely on visual perception for optimal functionality. Vehicles that drive themselves, robots, and surveillance systems are all included in this category. The field of deep learning has already made significant progress despite the fact that it is still relatively new. The performance of more recent models in computer vision, which make use of deep architectures such as CNNs, is superior to that of its predecessors. With the assistance of deep learning, pedestrian detection is made a great deal less difficult. An examination of the current state of the art in research and accessible methods is presented in this article with the purpose of assisting in the identification of pedestrians. This article takes a look at a few new models that are considered to be pioneers in their respective fields. Among the models that fall under this category are CNNs that are built on regions. Some examples of these CNNs are R-CNN, Fast R-CNN, and Faster R-CNN. Transformer-based versions and single-shot detectors (SSD, YOLO) are two other kinds of detectors. Methods for preparing data for processing, methods for adding new data, and methods for adjusting processed data to include non-maximum suppression are all covered in this research. Moreover, this study also includes methods for modifying processed data. During the presentation, each of these subjects will be discussed in further depth. It is possible to get significant insights into the operation of diverse systems by comparing the performance of these systems on various criteria such as recall, average miss rate, and accuracy. In conclusion, the purpose of this study was to investigate the current status of deep learning pedestrian detection systems in all of its exquisiteness.