The Vehicle Detection And Vehicle Model Recognition Based On Computer Vision

With the the rapidly increase in the number of vehicles,the demand of intelligent transportation and vehicle management is increasingly urgent.As a result,the technology of vehicle detection and identification based on computer vision is of important value and practical significance.In this paper,we have detected rear vehicles on the vehicle-loaded video.on the daytime,we use the Adaboost cascade detector to detect the vehicles.we have achieved the HAAR,LBP and HOG features.And,we train the model with second times based on the active learning to improve the performance of the model.Than,the sliding window strategy for different vehicle sizes is designed when detecting which greatly reduces the detection time.Also,we remove some false positives according to the continuity of video.In the nighttime,the method used on the daytime is not available.The tail lamp is the obvious characteristic of the rear vehicles at night.So,we detect the vehicles by detecting the vehicle taillights.Firstly,according to the characteristics of color and brightness,the region of red light is obtained.And then find pixels with extremely bright within the red light area to locate the red light source.Finally,do the pair matching operation.In this way,the rear lights are detected and the vehicles are detected.Also,vehicle Model Recognition is researched in this paper.Firstly,region of interest is detected by using the cascade classifier of Adaboost algorithm.Than,the popular classifiers,SVM and ELM,with SIFT,HOG and LBP,are implemented and compared.Importantly,an algorithm based on deep learning is proposed.Convolutional neural networks,an important member of deep learning,shows excellent performance especially in the field of images.We design a convolutional neural network with double supervised signals and find that its performance is greatly improved compared with traditional machine learning methods.Finally,we try to combine deep learning and traditional machine learning.We find that use the support vector machine to identify the features extracted from the last layer of the convolutional neural network,the recognition rate is increased by about 2%.