Deep Convolutional Neural Network-based Object Detection Methods With Applications To Autonomous Vehicle

ii Object detection is a very important content in vision system of intelligent vehicle.At present,the vision system of intelligent vehicle mainly take pedestrian and the vehicle as the research object.The pedestrian and vehicle detection algorithm is mainly based on the traditional methods which combined geometric features with the classifiers.These methods have already gaining a certain researching achievement,but under the actual situation which means the presence of pedestrians or vehicles could be blocked or deformed,the detection result is not able to meet the requirements.These negative factors could be overcome by the deep convolutional neural network algorithm.However,there are some weakness with the deep convolutional neural network algorithm,such as training and testing speed are slowly.In this paper,the pedestrian and vehicle detection in vision system of intelligent vehicle and its difficulties mentioned above have been researched.And then,the main study tasks and contributions of this thesis is given as follows.Aiming at the object detection task in the vision system of intelligent vehicles,the feasibility of deep convolutional neural network is explored.We propose an improved Fast-Rcnn algorithm based on Edgebox algorithm.The algorithm reference to human visual system saliency mechanism,using the object edge information structure pedestrian and vehicle targets potential candidate areas.On this basis,it combine with Fast-Rcnn to achieve a fast and accurate object detection algorithm.In the experiment,we compare the improved Fast-Rcnn algorithm and original Fast-Rcnn algorithm performance in Pascal VOC2007 object detection database.Finally,we use the image data collected in real autonomous vehicle experiment to verify the validity of the improved Fast-Rcnn algorithm.We propose a combining classification learning algorithm using improved Fast-Rcnn and OS-ELM(Extreme Learning Machine for online sequential data,OS-ELM),which can solve the slow convergence problem and the low accuracy problem in the Fast-Rcnn classifier.OS-ELM is a continuous,batch data input incremental learning algorithm.By combining it with the improved Fast-Rcnn algorithm,the accuracy of object detection gets further improved.In the experiment,we tested the performance of the combining classification learning algorithm in Pascal VOC2007 object detection database.And the results verify the validity of the algorithm.In this paper,a pedestrian detection database based on the real campus road environment is constructed.The database contains 4716 pictures includes 1179 manually labeled images.In the last Experiment,the combining classification learning algorithm using improved Fast-Rcnn and OS-ELM and the current mainstream ACF algorithm were compared.And the results also verify the validity of the algorithm.