Real-Time Multi-Target Detection Besed On Deep Neural Network

Target detection is a classic problem in the field of computer vision.It integrates knowledge in the fields of image processing,pattern recognition,and artificial intelligence.Compared with image recognition and image classification,the task of target positioning and target classification needs to be solved.In recent years,the rapid development of deep learning has greatly stimulated research in the field of target detection algorithms.However,the real-time nature of target detection has always been a major problem.The YOLO algorithms do a very good job in the real-time detection of the target,and the fastest version of and YOLO can detect objects up to 150 FPS.However,the increase in speed accounts for a serious decline in the detection accuracy.On the other hand,Faster R-CNN,the best of the R-CNN target detection series,performed excellently in terms of detection accuracy,and the average detection accuracy of 20 targets on the PASCAL VOC data sets reached 73.2 m AP,but in terms of real-time performance.However,it has not achieved satisfactory results.In response to the above issues,this article mainly carried out the following research:The final performance and effectiveness of the model depends not only on the structure of the model and the training techniques,but also on the quality of the data sets.In the field of target detection,there are many high quality public data sets,such as the PASCAL VOC data sets.However,most public data sets only contain static images,rather than video data sets.To achieve a real-time multi-object detection system,this article made part of my own data sets,mainly contains some video data sets in various scenarios.The target objects identified in this article are common five types of targets:pedestrians,cars,buses,bicycles(including motorcycles),and animals(cats and dogs only).And use the Label Img annotation tool to label the home-made data sets.The format of the annotation is the same as that of the PASCAL VOC data sets.Both are xml files.The existing video detection algorithms are mainly for image detection,and for videos is just treating each frame as a separate one-by-one image,ignoring that there is huge relevance between the upper and lower frames in the video data.This article considers the use of contextual correlation in video data,combined with Kalmanfiltering,sets a filter for each target detected,predicts its position in the next frame,and pay more attention to its predicted position areas,which can greatly improve The detection speed of the model and the Kalman filter tracking algorithm have a certain degree of anti-jamming effect on the occlusion problem of the target.For the target detection,the last step is to predict the probability that certain regions contain specific goals.A decision is made based on a fixed threshold.If the prediction probability is higher than this threshold,it is considered as a real target.If the prediction probability is less than this threshold,it is determined as a false target.However,there are certain problems with fixed thresholds.For situations where there are large differences in the scene or there are many different sizes of the target in the image,it is difficult for the fixed threshold to distinguish the high-probability false target from the low-probability real target.To solve this problem,this paper uses an adaptive threshold method to replace the traditional fixed threshold.Under the above improvements,the final experimental results of this model reached a video detection rate of 26 FPS.The m AP on the self-made data sets in this paper is 79.96.