Research On Human Target Detection And Tracking Technology Based On Video

With the development of modern image acquisition technology and digital video technology,more and more videos such as traffic monitoring,Internet video,and public security system video are needed to be processed in real time.It is very imminently to capture valuable information and ignore irrelevant information in video quickly and accurately.Strengthening object detection and tracking technology is very important for solving this problem.In recent years,research on object detection and tracking algorithms based on deep learning is so hot that the mature algorithms can not only free up a large number of human and material resources,but also bring benefits in accuracy and processing speed.Although the current technology can realize multiple targets realtime detection,most tracking algorithms are based on single targets.What's more,targets are often mistaken or lost due to occlusion problems during tracking.This study combines with deep learning technology,draws on some methods,ideas and model structures in recent years to improve the accuracy of detection.Analyzes the contribution of improved detection performance to tracking technology.Builds an association system for detection and tracking.A multi-target video detection and tracking algorithm was designed and implemented.In the end we yield a lightweightportable model for real-time detection and realize robust tracking of multiple human targets.In the detection part of the system,this paper completed a lightweight detection network based on multi-scale feature fusion of deep separable convolution structure for the real-time requirements of video processing.Added a loss function to alleviate overlapping problems of detection targets.Compared with the traditional convolutional neural network,this network greatly reduces the calculation amount and speeds up the detection speed.Greatly reduces the volume of the model and improves the recall rate of the detection target.So it's more suitable for combination with the tracking system.In the tracking part of the system,it combines with the fine-tuned detection model,Kalman filter principle,the association algorithm between frames and state transformation conditions.Apparent information matching degree and the motion information matching degree constraint are also added to realize the robust tracking of multiple human targets based on video scene.Experiments show that the video-based human target detection and tracking system implemented in this paper can be operated on computer real time.The repulsion loss function can better resist the interference caused by the target overlap.The depth separable convolution structure greatly reduces the volume of the model.When the target occlusion phenomenon occurs in the tracking scene,the designed association algorithm and the addition of two constraint mechanisms enable the occluded target to recover the tracking state after going out of occlusion.