Design And Implementation Of Multi-Objective Detection System Based On Deep Learning

With the popularity of mobile phone camera technology and the development of photographic equipment,a large amount of video and image data has been accumulated,which contains copious knowledge waiting for us to mine.As the development of artificial intelligence provides convenience for our life,we want to use the machine to do more tasks intelligently.So,machines should obtain more people's functions,such as vision,movement and language understanding.Vision plays an important role in all the perceptual perceptions of human beings.If computer obtains the visual function of human beings,the intelligence of computers will be greatly improved.Object detection is the core technology of computer vision,which is to frame the target we focus on from an image and mark the category of it.That is,it mainly includes two tasks,classification and positioning.Due to its fast detection speed and high detection accuracy,the multi-object detection technology based on deep learning has become the mainstream of current research areas,and more and more new algorithms have been proposed.In this paper,we analyze the advantages of the current popular multi-object detection technology based on deep learning through the development of target detection technology.The representative algorithms of multi-object detection based on deep learning are Faster R-CNN,Yolo and SSD.We analyzes the basic principles of these three algorithms and introduces their respective advantages and disadvantages according to the principle.Combing with our experimental needs,we compare and conclude that the Faster R-CNN algorithm is more in line with the needs of engineering tasks.In the training process of deep learning model,the quality and size of datasets and the settings of parameter are important factors affecting the training results.We collected the traffic image dataset by intercepting traffic video.Performing experiments based on Faster R-CNN,we experimentally analyzed the effects of different experiment settings on the results of model training,such as the scales of training dataset,different pre-training models,different training modes,different iterations,etc.The experimental results show that these factors have great influence on the mAP,accuracy and recall of the detection model,which can provide reference for the researchers to select the pre-training model and set the experimental parameters.