| Artificial intelligence technology has undergone significant changes in recent years,bringing convenience to people’s life and profoundly impacting target recognition technology.In addition,the commercialization process of autonomous driving is constantly unfolding,and the automatic driving of vehicles can be realized through multi-sensor fusion technology.Under this circumstance,this thesis focuses on the recognition and perception of multiple moving targets based on information fusion.The main research contributions and innovations are as follows:(1)To solve the problems of holes and noises in extracting the contours of moving targets with the inter-frame difference method,building a convolutional neural network is proposed to fix it.Firstly,using the inter-frame difference method to obtain the moving target,then using the constructed convolutional neural network to repair it and verifies the algorithm by recording data sets of different time periods,backgrounds,and targets.Experimental results show that the contour extraction of moving objects based on the convolutional neural network and the interframe difference method can repair the holes and noise problems that exist when the inter-frame difference method extracts the contours of moving objects.(2)Faster RCNN is used to complete the identification and detection of pedestrians and vehicles in the recorded data sets.By preprocessing the PASCAL VOC2007 data set,select the data sets containing images of vehicles and pedestrians;then use the processed data sets to train the model;finally,use the recorded data set for training and testing on the pre-trained model to achieve recording recognition and detection of targets in data sets.Experimental results show that Faster RCNN can identify pedestrians,vehicles and other targets in the recorded data set to a certain extent,but when the traffic scene is complex,or there are small targets or overlapping targets in the image,the detection effect of Faster RCNN is not very good.(3)Aiming at the problem of Faster RCNN unsatisfactory recognition effect in complex traffic scenes and inaccurate recognition of small targets and overlapping targets,the use of multi-sensor fusion technology is proposed to solve these problems.The target data obtained by the millimeter-wave radar and the target information returned by the RPN network are distributed and fused and then re-transmitted to the Ro I layer,thereby improving the accuracy of the target candidate frame and enhancing the recognition effect of Faster RCNN.Experimental results show that this method can improve the accuracy of Faster RCNN in recognition of complex,small targets and overlapping targets in traffic scenes.Still,it cannot recognize moving targets in traffic scenes.(4)Aiming to recognize moving targets in traffic scenes,it is proposed to use moving target extraction technology to obtain moving target information in images and then enhance Faster RCNN to recognize and detect moving targets in traffic scenes on information fusion technology.The experimental results show that the combination of moving target extraction and information fusion technology can improve the recognition accuracy of Faster RCNN and realize the recognition of moving targets in traffic scenes. |
