Computer Vision Based Pedestrian Detection

The pedestrian detection task is one of the important tasks in the field of computer vision.It integrates the knowledge of image processing,pattern recognition and machine learning to detect pedestrians visually.Through the research on pedestrian detection,pedestrian detection can be applied to many practical scenarios.There are three main ways to implement pedestrian detection:traditional image processing-based methods,machine learning-based methods,and deep learning-based methods.The focus of this article is on deep learning-based methods,starting from both the backbone network and the detection framework,to improve the effectiveness of pedestrian detection algorithms while ensuring speed as much as possible.The work of the thesis is mainly reflected in the following parts:(1)Aiming at the problem of insufficient feature extraction capabilities of light-weight backbone networks,this paper proposes a low-parameter,high-performance,and highly versatile substructure called Pooling Block.Pooling Block can be combined with almost all backbone networks to improve the feature extraction performance of the backbone network without affecting the running speed.(2)Aiming at the three problems of candidate frame generation strategy,convolutional layer and loss function of border regression in FCOS algorithm,this paper proposes candidate frame generation strategies called center and random-center strategy,deformable convolution and GIoU Loss to improve FCOS.Experiments show that these three improved algorithms greatly improve the results of pedestrian detection.(3)A SGDR learning rate strategy with a warm-restart is proposed,which improves the model's convergence results without affecting the training and detection speed.Various improved algorithms mentioned above have been tested on the Caltech Pedestrian dataset and the Citypersons dataset.The experimental results show that all the improved algorithms proposed in this paper improve the results of pedestrian detection,and at the same time can guarantee higher FPS.