Research On Human-vehicle Detection Method Based On Cascade CNN

With the rapid economic development of our country,the road traffic environment is increasingly improved and the car ownership keeps increasing.However,the continuous increasing in the number of cars also leads to traffic accidents frequent.Every year,the total number of casualties and property losses caused by traffic accidents is alarming.The incidence of traffic accidents,which has been increasing year after year,poses a serious threat to the lives and property of the people and makes negative impacts for the harmonious society building.Vehicles and pedestrians are the main participants in the traffic system and also as major victims in traffic accidents.Therefore,the research of real-time and accurate detection for pedestrian and vehicle has a major significance.The Network which cascades Initial Detection Network(IDN)and Accurate Classification Network(ACN)is designed to detect pedestrian and vehicle,according to the characteristics that the different categories(such as pedestrians and vehicles)differ greatly from one object to another,while the difference between the objects,belong to the vehicle category(such as bus,truck and so on)is relatively small.Firstly,the feature extraction parts of the IDN are constructed by transfer learning,and obvious differences among the objects of pedestrians and vehicles are detected according to the structural characteristics of IDN.The detection results of the vehicles are sent to the ACN to finish the subsequent calculations,in order to improve the detection accuracy efficiency by excluding non-target interference and reducing the amount of calculations.Then,by changing the form of the loss calculation function in the detection network,the ability of ACN to extract and learn the difference of feature between the objects,with small sample distance between each other,is enhanced.Finally,by cascading the two networks the final network ICAD is constructed to complete the detection task.The ICAD method can significantly reduce the number of parameters in the detection process and improve the detection efficiency under the premise of ensuring the detection effect.The test results of open source datasets created by public and the experimental platform show that this method can effectively improve the detection speed with satisfying the detection checkup rate and reduce the false detections in a certain extent.