| Ferry boats play an important role in public transportation,and their reasonable dispatch is of great significance to meet people’s needs and improve resource utilization.Multi-object detection technology can provide important references for the reasonable dispatch of ferry boat resources by detecting the various vehicles which get on or get off the ferry boats.However,due to the complexity of the ferry boat scene,four types of objects,including person,car,truck and bus,need to be detected.Therefore,the realization of a detection algorithm with high detection accuracy is of great significance for the multi-object detection in ferry boat scene.Faster R-CNN is a target detection algorithm with high detection accuracy in the field of target detection in recent years.It uses the multi-layer model to learn abstract representations of data and can discover complex structures in lots of images.The way to use this algorithm to detect multiple targets in ferry scene is the research content of this thesis.The research work is as follows:By deeply researching on the the detection process of Faster R-CNN algorithm and combined with ferry scene,RPN and classification regression network are set up and the training method is determined.Besides,ResNet101,with higher detection accuracy on the built data set,is selected as feature extraction network.After training and testing of the detection model,the experimental results show that the Faster R-CNN algorithm may miss detection of small size targets in ferry scene,and the detection accuracy is low.For the problem of Faster R-CNN algorithm in detecting small-size target class in ferry scene,this paper proposes an improved method based on fusion features on the basis of this algorithm.Through the fusion,the high-resolution features of the low-layer feature map can be combined with the high-semantic features of the high-layer feature map.When generating the RoIs,the corresponding fusion features in the feature set are selected according to the size of the suggested area,and the classification and regression are performed on this basis.In this paper,the improved algorithm is trained and tested on the self-built data set.The experimental results show that the improved algorithm effectively improves the detection accuracy of small size objects. |
