| With the prosperity of computer vision,deep learning-based object detection technology has become one of the mainstream approaches to solve the problem of autonomous driving,and the performance of network architecture is in a crucial position in object detection.The autonomous driving requires a high level of neural network in terms of accuracy,speed,and stability,etc.The traditional approach based on manual design of the network is a little weak because it not only requires researchers to have sufficient reserves of theoretical knowledge,but also requires continuous "trial and error" for structural improvement to come up with a suitable network model.As a result,people began to focus on neural architecture search,which requires only a simple design of the search space and can be automated to complete the construction of the network in a specific scenario,saving a lot of human and material resources.However,due to its search efficiency,neural architecture search does not perform well in object detection,and the long search time hinders its advancement to industrial scenarios such as autonomous driving.To address this problem,this work revolves around how to accelerate the object detection search model.As the research progresses,it is found that the gradient descent search strategy in the classification domain can significantly reduce the search time,and if the " collapse" problem of this algorithm can be solved,it can be migrated to the object detection domain,thus significantly improving the search efficiency for the object detection problem.In summary,the work of this thesis is divided into the following three parts.1.In order to solve the " collapse" problem of the gradient search model,this thesis analyzes the existing experimental results and concludes that the specific cause of the " collapse" is the anomalous change of the skip connection weights.In this thesis,we add dynamic regular coefficients to the skip connections to balance the competition among operations in the search space and provide a stable classification search model for subsequent work.Compared with the baseline model,the accuracy of DR-DARTS improves 0.76% on the CIFAR-10 dataset and 1.82% on the Image Net dataset,while the number of parameters is only 3.1M.2.In this thesis,we propose MOD-NAS,a migration-based object detection network search model,to study the commonality of classification and object detection problems,and propose a model migration-based and weight sharing approach for object detection backbone network search.The work is evaluated in the publicly COCO dataset,and it takes only 170 hours to complete a single search,which saves about 44%of the search time compared with the same model in the same period,while the model m AP index reaches 39.4.3.Object detection practical task measurement.In this thesis,we simulate the object detection task in a real autonomous driving scenario using an intelligent vehicle and a manually built track.Using the efficient search model proposed above,we custom-build a lightweight object detection network that meets both accuracy and inference speed requirements,locate and detect obstacles and markers on the track in real time,and combine it with the written intelligent vehicle control program to complete the autonomous driving task of the intelligent vehicle. |
