Study On Night Vehicle Target Detection Method Based On Convolutional Neural Network

With the rapid development of urbanization in China,the waste of urban resources has always been an urgent problem to be solved.For example,when the city enters the late night,the traffic flow on the street becomes less,and sometimes there is no vehicle at all,but the street lights will always work,which will cause huge waste of energy.It is a feasible method to control the brightness of street lamps by traffic flow.In addition,we can make full use of video obtained by the street camera to detect the traffic flow through video analysis.This method will not have additional hulan and financial burden and has a strong practical significance.With the rise of deep learning,it has become a trend to use convolutional neural networks to process image tasks.This paper introduces the application of deep learning network in object detection task in detail,and designs and optimizes the scheme that can be used for vehicle detection at night.Firstly,this paper introduces the traditional BP neural network and convolutional neural network.We compare different object detection networks and analyze their advantages and disadvantages.Considering the requirements of the research topic in this paper,the R-FCN object detection network was finally selected,which introduced translational variation characteristics and was very suitable for object detection tasks.Secondly,in order to improve the speed and accuracy of network detection,AssignParallel mechanism and Soft-NMS algorithm are applied to R-FCN network.AssignParallel mechanism can compress the feature extraction part of R-FCN network.On the premise of ensuring accuracy,AssignParallel reduces the physical memory occupied by the model and improves the network speed.Soft-NMS algorithm can solve the problem of mutual concealment of objects and further improve the detection accuracy of the network.At last,this paper compared the detection results of R-FCN network and Faster RCNN network at different time periods at night.At the same time,we conducted model compression experiments on R-FCN network and Faster-RCNN network and compared the accuracy with the original network to verify the feasibility of AssignParallel mechanism.In addition,we also compared the Soft-NMS algorithm with different parameters.Compared with the traditional night vehicle detection algorithm,this algorithm has higher detection accuracy,less physical resources and better practicability.