| As the urbanization accelerates,people crowd into large cities,increasing the risk of accidents in crowded places.Traditional video surveillance technology is difficult to adapt to the current requirements,with the rapid development of deep learning technology,intelligent video surveillance technology began to rise.Intelligent video surveillance technology is an important tool in the pedestrian safety research field.Through the real-time processing of video information,it can respond to unexpected situations in a timely manner and ensure pedestrian safety to the greatest extent.Compared with the intelligent monitoring of vehicles,pedestrians are more difficult to supervise than vehicles due to their changeable routes,large density,occlusion,while the existing pedestrian tracking models still have room for optimization in accuracy and speed.The main contents of this paper are divided into the following 5 parts:1.Introduce current mainstream object detection models: Two-stage,One-stage and Anchor-Free;object tracking models: SDE and JDE.Introduce the pedestrian flow theory as the basis for pedestrian traffic evaluation.2.Introduce the theoretical basis of object detection and tracking model based on deep learning--convolutional neural networks and related theories.3.Compare the three mainstream object detection models and two target tracking models,summarize the advantages and disadvantages of the current model,propose a pedestrian target tracking model based on Center Net-JDE,and introduce the model framework and loss function.4.According to the characteristics of pedestrian obstruction and small targets,a pedestrian dataset MPD is made,and used to train detection model;the tracking model is trained by using the public dataset MOT-16.Experiments show that the detection model Center Net is 5.9% more accurate than faster than faster than Faster-R-CNN’s AP50 and 41.2% faster than YOLOV3.In terms of tracking models,the Center Net-JDE model in this paper improves the MOTA accuracy of the baseline model YOLOV3-JDE by 8.9%.5.Collect pedestrian videos of a subway station in Beijing for empirical analysis.The Center Net-JDE model is used to collect pedestrian traffic parameters in the video,and the data is analyzed and evaluated. |
