Study On Vehicle Detection And Classification In Emergency Lane Based On Logistic Regression And CNN

Vehicles occupying emergency lanes have seriously affected the normal use of emergency lanes.At present,the screening of captured images of vehicles occupying emergency lanes mainly depends on manual processing.A vehicle detection and identification method for the captured image of emergency lanes is proposed in this paper,which can realize intelligent detection of vehicles in the emergency lane and automatic judgment of the type of the vehicle.This paper introduces the theory of logistic regression and proposes a vehicle detection method based on logistic regression;based on the feature abstraction and automatic learning ability of convolutional network,the vehicle identification algorithm based on convolutional network is studied;analyzing the contribution of convolution kernels in convolutional networks and the convolution kernel cutting method based on the sum of convolution kernel variance and its L1 norm is proposed to compresses and improves the CNN model.Using the HOG feature of the training set to train the logistic regression classifier,searching the target area by the sliding window mechanism,extracting the HOG feature of the sliding window and sending to the classifier,and the detection result is output by the non-maximal suppression technology;the CNN model is trained using the vehicle type training set produced in this paper,and cutting the convolutional kernel of the CNN model based on the proposed method in the paper,then retraining the after cutting CNN model with the training set so as to achieve optimal performance,finally putting the vehicle detection result into the CNN model to obtain the vehicle type recognition result.The main research content is as follows:1.Vehicle detectionAnalyze the classification principle and parameter solving method of logistic regression,and on the basis of the HOG features of vehicles,a vehicle detection method based on logistic regression is proposed.2.Vehicle classificationAnalyze the structural characteristics and parameters of the convolutional network and design a deep convolution network model for vehicle type identification.3.Model compressionAnalyze the contributing factors of the convolutional kernel for feature extracting in the convolution layer,such as the average zero proportion,L1 norm,etc.And on basis of this,a convolution kernel cutting method based on the evaluation indicator of the sum the convolution kernel variance and its L1 norm is proposed,which can achieve model compression.The experimental results show that the detection rate of the proposed vehicle detection method on the captured emergency lane data set is 97%,which is 0.9% higher than SVM and 2.3% higher than Adaboost,and the average detection time is about 30%-40% shorter than that of SVM and Adaboost.Compared with SVM and Adaboost,the logistic regression model is simpler and less complex;besides,the accuracy of the vehicle identification algorithm based on CNN is up to 99.2%,which is about 8% higher than that of manual extracting feature methods;retraining the CNN model after cutting the convolution kernel,the results show that the parameters of the model have been reduced by 50.7%,and the model computation time has dropped significantly.However,the accuracy rate can still be maintained at 99.1%,which verifies the effectiveness of the cutting method and also proves the practicality of the CNN model.