| Stereo garage is an important way to solve the problem of urban parking difficulty,which alleviates the shortage of parking land through rational use of space resources.However,stereo garage not only brings convenience,but also brings a lot of security problems,especially unattended stereo garage.How to effectively ensure the safety of pedestrians in the parking area has always been the key problem to be solved in unattended stereo garage.With the support of the key R&D project in Jiangsu Province,this paper researched the pedestrian detection method in garage area based on deep learning,which can quickly and accurately complete the pedestrian detection in the complex scene of practical application,and provide solid technical support for the development of unattended stereo garage.Firstly,according to the use of stereo garage,the existing problems and challenges of garage pedestrian detection system are described,and the overall scheme of the system is designed.Secondly,in this paper,a garage pedestrian detection method based on improved convolutional neural network is researched and implemented.In view of the problem that the surveillance video contains a large number of non pedestrian idle segments and wastes computing resources,the inter-frame difference method is adopted to realize the purpose that the pedestrian detection model only deals with non idle segments.Aiming at the problem of people being occluded by scene objects,a RandomMix data augmentation strategy is designed to effectively improve the pedestrian detection ability of the model in occluded scenes.For different illumination scenes,this paper designs a pedestrian detection method based on day and night pedestrian detection model separation and an illumination compensation algorithm based on Multi-Exposure Fusion to enhance the adaptability of the model to different illumination scenes.Thirdly,in order to improve the real-time performance of pedestrian detection model,a lightweight method of garage pedestrian detection model is proposed.A branch fusion network is designed and added into the backbone network.The feature of each stage in the network is separated by the branch structure to reduce the forward inference computations.In addition,the features of multiple branches are fused at the corresponding nodes to achieve complementary feature information and enhance the learning ability of the network.Experimental results show that the lightweight method can reduce the network size while slightly reducing the detection accuracy,improve the speed of pedestrian detection,and shorten the time delay of pedestrian detection.Finally,in order to further improve the accuracy of pedestrian detection,a garage pedestrian detection method based on Transformer is researched and implemented.In this paper,Transformer is introduced into the garage pedestrian detection model in two different ways.The first way is to integrate Transformer and CNN.Transformer is regarded as a powerful supplement to CNN,and the self-attention mechanism is used to capture the full range of related dependencies of each local feature in the convolutional neural network.The second approach takes Transformer as the basic building module,constructs a complete attention mechanism model,and uses the self-attention mechanism to extract and process features.Experimental results show that the above two garage pedestrian detection methods based on Transformer have achieved higher detection accuracy than the convolution neural network based garage pedestrian detection method. |
