Study On Multi-Obstacles Detection And Recognition Under The Car Driving Environment Based On Convolutional Neural Network

With the car ownership increasing,traffic accidents pose a serious threat to peoples’ s lives and property.Failing to timely and accurately detect the obstacles around the vehicle is the main cause of frequent traffic accidents in a complex and changeable driving environment.In the era of intelligent driving,we propose some new requirements about the obstacle perception technology under the car driving environment for guaranteeing travel safety.The new perception system should be real-time with a high recognition rate.Drivers get 80% of the information from visual perception.Therefore,studying the visual-based obstacle perception technology contains great potential for application.Traditional obstacles detection and recognition methods are difficult to meet the requirements of the smart car and unmanned vehicles in terms of recognition accuracy,reliability and universality.Moreover,traditional detection equipments are expensive.The main work of our goal is to solve the above problems.To design an obstacle detection and recognition system with high recognition accuracy and universality,focus on the deep learning algorithm and apply the mehod based on convolutional neural network to detect and recognize the obstacles under the car driving environment.The main contributions of our work are as follows:Study the algorithm of obstacle region extraction,analyze four types of ROI extraction methods,including image basic feature-based,visual attention model-based,human-computer interaction-based and object-based.On this base,in consideration of the huge amount of vehicle video data,an improved method that using Otsu algorithm based on regional growth algorithm combined with morphological operations is proposed,which can extract obstacle areas automatically.The method transforms the vehicular image from RGB color space to Lab color space,uses pixel gray level characteristics and the largest interclass variance to divide the obstacles and background,applys morphological operations to eliminate images redundant objects,fill the area void,smooth the border,uses the regular graphics to extract the obstacle part from the car image,and creates the obstacle dataset.Construct a deep convolutional neural network for detecting and recognizing multiple types of obstacles under the car driving environment.We build a 15-layer network for obstacle detecting and obstacle area recommending based on AlexNet and the Caffe framework from the Deep Learn Toolbox.The network includes a five-layer convolution layer and a two-layer pooling layer for obstacle feature extraction,a fourlayer RRN network for generating an obstacle recommendation area,a layer of ROI Pooling acting on each of the regions of interest to generate a fixed size map of feature,a two-layer fully connected layer for reducing data dimension and further extracting the feature of object,an output layer setting up five neurons,that is,there are five categories of obstacles detected and recognized.Data from the obstacle datasets and PASCAL VOC 2007 and PASCAL VOC 2012 databases are used for training and testing the deep convolutional neural network.To reduce network complexity and solve the problem of over-fitting,the network optimization strategies are used to optimize the deep convolutional neural network.To observe the rationality of network structure,free parameters of the network training process are visualized and analyzed.To horizontal inhibit the output neurons,increase the local response normalization layer,which can improve the generalization of the network and target detection accuracy.Study network over-fitting problem,use Dropout and DropConnect methods to solve the network over-fitting problem.The improved effect of both two kinds of de-fitting methods on the recognition of target are analyzed experimentally.Collect car videos under the real environment,and use them to test and verify the performance of the proposed obstacles detection and recognition method.The results of the experiment show that our proposed method can carry out high precision and real time detection and recognition about various types of obstacles such as car,pedestrian and motorcycle under the car driving environment.Compared to the traditional Gaussian mixture model combined with Kalman filter obstacle detection method,the obstacle detection categories are completed and the detection accuracy are improved.The algorithm proposed in this thesis is more universal,the cost of the test equipment is lower,that means,our proposed method possesses good theoretical value and practical value.