Research On Interpretability Of Deep Learning Models In Unmanned Driving Scenarios

Deep learning model has extremely superior performance in many scientific research fields,and many researchers begin to apply deep learning models to the field of unmanned driving,which makes important breakthroughs in some key technologies for unmanned driving.But deep learning models are more like a "black box" whose internal decisions and behaviors are agnostic to humans.Unknowable and unexplainable model also means danger,which triggers a crisis of user trust in deep learning models and driverless technology.Therefore,it becomes more and more important to explain the deep learning models applied in the field of autonomous driving.Based on this problem,this thesis first builds three driving scenes in the Carla simulation environment as a platform for depth learning model training and interpretability algorithm interpretation and analysis.This thesis plans the route in the driving scene in advance,controls the vehicle to travel according to the planned route through the horizontal control algorithm and the vertical control algorithm,and collects the required picture data and text data in this process.After that,a deep learning model with convolutional neural network as the core is built,and the data collected in advance are used as the training set.A convergent unmanned driving deep learning model is obtained after training,that is,the model to be explained.Finally,two local interpretability algorithms based on perturbation technology are proposed to explain the convolutional layer and fully connected layer of the deep learning model respectively,highlighting the areas that play an important role in the model decisionmaking process:(1)The first algorithm is a "local interpretability algorithm based on image perturbation",which perturbs the input image(input data),and displays the change of the last convolution layer output in the original image in the form of a saliency map.(2)The second algorithm is a "local interpretable algorithm based on the disturbance of the characteristic map".It disturbs the characteristic map(intermediate variable)of the last convolution layer,takes the change of the output of the full connection layer as the weight,multiplies it with the corresponding characteristic map,adds it to the original map,and displays it in the form of a saliency map in the original map.In view of the above work content,this thesis strictly follows the standards of Road Traffic Signs and Markings(GB-5768)during the construction of driving scenes to make driving scenes consistent with real scenes.Whether the trained model is converged and stable is judged by the loss function curve of the model in the training process of each scene.And the performance of the model is evaluated through the Benchmark tool that comes with the Carla environment.The results show that the model built in this thesis becomes convergent and stable after 300 iterations of training,and performs well in various driving scenarios,and is a qualified model.The results show that the model built in this paper becomes convergent and stable after 300 times of iterative training,and performs well in each driving scene.It is a qualified model.Then,the decision and behavior of the model in each scene are explained through "Local interpretability algorithm based on image perturbation" and "Local interpretability algorithm based on feature map perturbation",and compared with existing algorithms.The simulation results show that the convolution part obtains the most important regions,and the full connection layer will select a part of these important regions as the basis for decision-making.At the same time,in the process of model training,the interpretative algorithm is used to observe the convergence of the model’s focus area,which further verifies the rationality of the interpretable algorithm proposed in this thesis.