Research On Vehicle Detection In Traffic Surveillance Video

With the gradual establishment of intelligent transportation system,the demand of vehicle detection technology in traffic monitoring video continues to expand.In this paper,the current problems of detecting vehicles in video and the existing vehicle detection methods are throughly studied.Then this paper presents a novel vehicle detection method combining Vi Be algorithm and Faster R-CNN.The main contents of this paper are summarized as follows:1.To detect vehicles in video,this paper propose a novel vehicle detection framework which combines the moving target detection algorithm(ViBe)and object detection method(Faster R-CNN)in static image based on deep learning.This framework is designed and implemented to solve two problems:1)Conventional moving target detection methods do not perform well in complex traffic conditions such as light mutation,nonstable glint,vehicle occlusion.This paper applies Faster R-CNN to learn strong features from large number of training data under complex conditions.Thus being able to detect the vehicle in complex traffic conditions accurately.2)In order to further reduce Faster R-CNN calculation time and avoid the occurrence of false positives,this paper applies ViBe algorithm to quickly extract the moving target area,which can greatly reduce the amount of computation and bring the false positive rate down.2.In order to achieve the efficient and uniform fusion of the ViBe algorithm and Faster R-CNN,this paper improves both ViBe and Faster R-CNN in the following aspects:1)The detection result from Faster R-CNN is used to help initialize the ViBe background model.This initialization strategy can avoid the occurrence of the ‘ghost’.2)The detection result from previous frame is applied to guide the model update and foreground judgement strategy of ViBe,which enhanced our ViBe algorithm‘s robustness when dealing with complex situations.3)Features from different convolution layers are combined,so the classifier can access more levels of information and have stronger ability to detect the small vehicles in the video correctly.4)The temporal information is considered to filter the proposals extracted by RPN,which can reduce about 50% of the proposals while maintaining accuracy rate.5)A novel weighted temporal non-maximum suppression algorithm is proposed.The weighted temporal NMS models the temporal results to re-score the classification score and filter the result,which help our Faster R-CNN get the correct result in those complex scenes that it fails to detect the vehicles.3.This paper builds a vehicle dataset based on the surveillance video provided by the Guangdong transportation group.This dataset contains more than 10000 positive samples of vehicle.The method proposed by this paper is tested on this dataset.The experiment result shows that our method has both high accuracy and speed comparing to the existing method.