Vehicle Detection Based On Convolutional Neural Networks

In recent years,machine vision has been widely used for different applications.As a key technology of the Intelligent Transportation System,vehicle detection technology which aims to automatically detect vehicles from images captured by camera,has became a hot research area.However,previous vehicle detection algorithms are often sensitive to illumination,scale variance,occlusions and complex background,this paper introduces the Convolutional Neural Networks(CNN)to address these problems which pose the vehicle detection as a classification problem.Moreover,this paper also presents an Discriminant Convolutional Neural Networks(D-CNN)algorithm that combine Softmax classifier and Linear Discriminant Analysis theory.The main contents are abstracted as following:1.This paper proposes a improved CNN algorithm that combined with the Linear Discriminant Analysis theory.Due to the loss function and test accuracy rate curves shock seriously while the CNN model is trained,this paper proposes a D-CNN algorithm which introduce Linear Discriminant Analysis theory into the Softmax classifier,and constructs the new hypothesis function and loss function.The D-CNN algorithmcan improve the feature recognizable via reducing the distance of inner-class samples and increase the distance of outer-class samples.The new algorithm can obviously smooth the loss function curve and improve the test accuracy rate through comparing the experiment result.2.This paper presents a vehicle type classification method via the improved CNN algorithm.The algorithm adopts a nine layer Convolutional Neural Networks model and uses the car,truck and coach images to train the model.The experimental results show that the D-CNN classification model can effectively improve the robustness to illumination,occlusions and complex background and test accuracy rate via comparing the test result to classical classification algorithm.3.This paper conducts the vehicle images detection by using the D-CNN classification model.This paper extracts about 2000 candidate area images via Hierarchical Grouping Algorithm and acquires the vehicle detection model through using the candidate area images to fine-tune the improved CNN model.Finally,the experimental obtains the vehicle location information via analyzing the classification model output category information.The experimental results show that the D-CNN classification model can effectively reduce the leak detection and error detection that because the illuminate,occlusions and complex background.