Research On Pedestrian Detection Algorithm In Railway Operation Lines

Railway is an important infrastructure in our country.With the continuous expansion of the construction scale of railway operation lines in China,the risk of accidents during construction is also increasing.At present,the safety management of construction site operations is mainly based on the low-efficiency civil air defense mode,which has some security risks,and it is urgent to introduce intelligent means of safety protection.Therefore,it is of great significance to study the pedestrian detection algorithm in the railway construction scene for the design of the intelligent security monitoring system and to ensure the construction safety along the railway.However,due to the complex environment of the construction site and the presence of many small-scale pedestrians in surveillance videos,pedestrian detection in construction scenes confronts great challenges.This thesis conducts an in-depth study on how to improve the performance of pedestrian detection in railway construction scenes.with the following main work accomplished.:(1)In view of the lack of railway construction scenes in the existing public datasets,a railway construction scene dataset is constructed,which contains several typical construction environments.In addition,according to the different characteristics of the scenes in the dataset,the dataset is divided into multi-scale pedestrian detection scenes with few occlusions and complex background detection scenes with serious occlusions.The design of the subsequent algorithm is carried out for the above two scenarios.(2)Aiming at the problem that small-scale pedestrians are difficult to detect in multiscale pedestrian detection scenes,a multi-scale pedestrian detection algorithm based on YOLOv5 is designed.First,a self-mimic learning method is introduced to make smallscale pedestrian feature information imitate large-scale pedestrian feature information,enhance the feature of small-scale pedestrians,and improve the problem that small-scale pedestrian feature information is weak and difficult to distinguish from the background.Second,the original structure is replaced by a bidirectional feature pyramid network,so that the network can autonomously learn the importance of each feature layer during feature fusion.Finally,the quality focal loss function is used to optimize the classification and confidence loss functions,so that the network can quickly focus on the learning of hard samples and positive samples during training,which improves the performance degradation caused by the imbalance of positive and negative samples during training.The experimental results show that,compared with the basic detection network,the multiscale pedestrian detection network designed in this thesis improves the detection accuracy of multi-scale pedestrians and reduces the miss rate.(3)Aiming at the problem of many occlusions in the complex background pedestrian detection scene,a pedestrian detection algorithm based on data enhancement and attention mechanism is designed.Applying the Mixup data augmentation strategy increases the adaptability of the model to complex environments.At the same time,a joint attention mechanism is constructed to make the model focus on the learning of pedestrian features from both the channel dimension and the spatial dimension to improve the network’s ability to detect occluded pedestrians.The experimental results show that,compared with the basic detection network,the network structure designed in this thesis is more suitable for pedestrian detection tasks in occlusion scene,which improves the accuracy of pedestrian detection and reduces the miss rate.