Research on Perception System of Automated Driving Based on the Monocular Vision - Vehicle Detection

— In order to build self-driving sensing system based on monocular vision, this paper studied the detection of vehicles in the driving environment based on HOG (Histogram of Oriented Gradient) features and SVM (Support Vector Machine) classifier. To solve the problem of low detection efficiency which will be caused by the higher dimensions of HoG feature space, a HoG-PCA (principal component analysis) feature based on the dimensionality reduction was proposed. After testing, it was proved that the feature after the PCA dimensionality reduction does not affect the detection accuracy. In the process of classification based on the SVM classifier, the distribution of data was unknown, so the kernel function of SVM was studied. At the same time, Adaboost (Adaptive Boosting) algorithm was used for the classifier comparison. Based on the ensemble learning method, a new classification model based on SVM classifier with different kernel functions and Adaboost algorithm was built. After testing, the new model can achieve a higher accuracy of 98.5%. Finally, the new model was used to detect the vehicles in the video and the deep learning model SSD (single shot multibox detector) was introduced to compare with this new model. The results show that, even compared to the latest object detection method based on the deep learning method, the traditional classification methods based on feature still have some reference value.