Research On Vehicle Detection Technology Inintelligent Driving Scene

With the rapid increase of car ownership and the complication of road traffic systems,the number of traffic accidents has remained high.To alleviate this situation,intelligent transportation technologies such as intelligent assisted driving and driverless driving have emerged.As the main body of the accident,vehicles are the targets that need to be detected in each system of intelligent transportation.The quality of vehicle detection is the key factor for the entire system to function.This paper focuses on the three difficulties of vehicle detection: nighttime vehicle detection,occluded vehicle detection,and small target vehicle detection.For nighttime vehicles,the traditional image processing technology is used to extract the lights for detection.For occluded vehicles,the traditional machine learning multi-feature cascade classifier is used for detection.For small target vehicles,a detection method based on deep learning convolutional neural networks is studied.The main research work of this paper is as follows:First,a nighttime vehicle detection method based on vehicle lights extraction is proposed.In order to adapt to the scene of different street light illumination at night,the region of interest(ROI)of image is gamma corrected to highlight the area of the vehicle lights.Aiming at the problem of unsatisfactory segmentation effect of Otsu algorithm,one-dimensional maximum entropy is used to improve Otsu algorithm.In order to eliminate the influence of residual interference terms such as reflected light,a priori knowledge such as the symmetry and aspect ratio of the lights is used to match the light pairs to detect vehicles at night.Experimental results show that the proposed method can achieve good detection results under various street light illumination intensities,and can detect lights from different directions,and is applicable to both headlights and taillights.Then,an occluded vehicle detection method based on multi-feature cascade classifier is proposed.Haar-like,complete local binary patterns(CLBP)and histogram of oriented gradient(HOG)are used to extract detailed edge,texture and contour features respectively.Then use principal component analysis(PCA)to reduce the dimensions of the three features,remove the redundant information,fuse the features,and send them to the classifier for detection.The adaptive boosting(Ada Boost)cascade classifier was selected to reduce the computation time,and in order to avoid overfitting,the support vector machine(SVM)algorithm was combined to improve the cascade classifier.If the false alarm rate of a certain level does not meet the preset conditions,replace the Ada Boost classifier with the SVM classifier.The positive samples required for classifier training are manually captured vehicle parts maps to accommodate various vehicle occlusion situations.The experimental results show that compared with the SVM classifier based on HOG features,the Ada Boost classifier based on Haar-like features,and the Ada Boost classifier based on multi-features,the proposed method can achieve better detection results,and can also achieve time real-time requirements.Finally,a small target vehicle detection method based on feature fusion SSD(Single Shot MultiBox Detector)algorithm is proposed.The single-step network SSD algorithm with obvious computing speed advantage is selected.The 6-layer feature maps extracted by the algorithm is completely independent,and the shallow feature map suitable for detecting small targets contains too little semantic information,so the detection effect on small targets is not ideal.In order to enhance the semantic information of shallow feature maps,a feature pyramid network(FPN)is used for layer-by-layer feature fusion for the 6-layer feature map,and then the fused 6-layer feature map is secondarily enhanced using a balanced semantic method,and finally vehicle detection is performed.The semantic balance is to average the semantic information of the 6-layer feature maps at one time,and then use the convolution operation to enhance the features.The obtained feature maps are fused with the 6-layer feature maps of FPN to achieve the secondary enhancement of the feature maps.Feature maps get the same amount of semantic information from other layers.Experimental results show that compared with YOLO(You Only Look Once)algorithm,ordinary SSD algorithm,and SSD-FPN algorithm,this method can achieve effective detection of small target vehicles while ensuring time efficiency,and improve the average accuracy of detection.