Point Processes For Vehicle Recognition And Extraction In Remotely Sensed Images

With the increasingly high resolution of remotely sensed image,the information about objects contained in an image is becoming more and more abundant.Consider a Vehicle,one of the most found objects in remotely sensed images,its quantity is huge,that is quite complicated to be identified manually.This paper focuses on vehicle recognition and extraction problem in remotely sensed images,and uses point processes to establish a probabilistic model for vehicle detection.The probabilistic model includes a prior model and a data term.It analyses the distribution of vehicles on space in prior model,then connects the model with the real image in data term,and finally uses Simulated Annealing and Reversible Jump Markov Chain Monte Carlo to get an optimized result of the model.The main work is summarized as follows:(1)A prior model that can reflect the relationship between vehicles is proposed.It uses a random point to mark the vehicle,and give it properties,like the width,height,and rotation angle,resulting in the geometric model of the vehicles.Vehicles could be expressed as rectangles in the image,and vehicle detection can be regarded as a random distribution of multiple rectangles on space.Vehicles should not be overlapping.When two rectangles intersect,we assign a zero-probability to forbid this phenomenon.The direction of the vehicles tends to be same.When it is found that a vehicle is not in the same direction,it gets a low probability to counter this phenomenon.(2)The similarity of matching template is introduced in the data term,and some improvement is achieved.Choose a vehicle from the image as the template,calculate the similarity between the template and the image,and establish data term through the similarity value.The direction of the vehicles is variable in the real world scenario,single direction template cannot detect such changes in direction of the vehicles.So we rotate our template within 360 degrees to adjust according to differently directed vehicles.Vehicles have also diverse colors in the images,so due to different colors,different color templates should be chosen to match with,therefore we conduct training samples of different vehicle colors to adapt to a variety of colors of vehicle.Finally,by using five marked vehicles images for evaluation of our algorithm,the precision can reach more than 99%,and the recall is more than 90%.The experimental results show that the prior model greatly improves the vehicle detection.The overlapping limitation of rectangle keeps the detected vehicles separated from each other,while the punishment of the rotation angle makes the vehicles keeping a consistent direction.