Multi-object Detection,tra Cking And Event Detection Based On Moving Cameras

In recent years,with the continuous development of computer vision technology,more and more computer vision related products have been put into our daily lives,and bringing us a lot of convenience in life.However,there are still many issues worth studying in vision,related to moving cameras such as drones and traffic surveillance.This thesis focuses on the three issues which are multi-target detection,tracking and event detection in videos captured by moving cameras.For object detection in UAV shooting videos,we analyze the proportion of small objects in videos and the difficulties for training nets of small object detection.And then improvements are made based on the ReitnaNet so that more small objects can be detected.At the same time,for the multi-target tracking,we propose a multi-target tracking strategy which can automatically identify irregular movement of drones.These methods were validated on the VisDrone dataset and achieved second place in the 2018 VisDrone Challenge.For the abnormal event detection problem of traffic monitoring video,we propose a detection method based on vehicle trajectory.An unsupervised method for extracting road masks is presented to improve the robustness of our method.With this method we achieved seventh place in the 2019 AI City Challenge.We also use the methods of object detection and multi-target tracking in the video to explore vehicle speed measurement tasks.In our method,the vanishing point of the image is used to calibrate the projection matrix of the camera,while the real-time detection and tracking strategy is used to extract the short-term trajectory of the vehicle to further calculate its instantaneous speed.