Research On Multi-vehicle Detection And Tracking Technology Based On Roadside Sensors

In recent years,auto driving technology has become a hot research field.We can extract the position and trajectory of the vehicle by detecting and tracking the vehicle using roadside sensor.These information will be broadcast to the unmanned vehcile,which could assist the ummanned vehcile to drive automatically.In actual road scene,there are various challenges in the detection and tracking of vehicles.On the one hand,uncontrollable factors such as weather change and light intensity affect the road environment at all times.On the other hand,different sizes of vehicles and the occlusion between vehicles also bring many difficulties for detection and tracking.In addition,there is also a challenge to ensure both the accuracy and timeliness of the system in practical applications.In this thesis,we propose a light-weight and fast vehicle detection model based on deep learning and a multiple object tracking method based on spatial relationship and color features.Based on the above method,we design and implement a prototype of a multi-vehicle detection and tracking system based on roadside sensors.In terms of vehicle detection model,we migrate the lightweight network MoblieNet to the Faster R-CNN object detection framework and optimize the RPN structure in order to meet the requirements of accuracy and timeliness at the same time.Through these methods,we greatly reduce the quantity of parameter and computation and improve the detection speed.In terms of object tracking,we calculate the matching probability of objects in continuous two frames by the spatial relationship and color feature.For the missed detection caused by occlusion,the KCF tracker is used for position prediction.Finally we achieve long-term tracking of multiple obj ects.We conduct some experiments in pratical scences.The experimental results show that our method has high multi-vehicle tracking accuracy and fast running speed,which meets the research expectations.