Research On Real-time Object Detection Algorithm Based On YOLO

The real-time object detection algorithm based on deep learning has become popular in research and application.The real-time object detection algorithm has been widely used in many fields such as intelligent transportation,intelligent monitoring,security,military,surgical medical and so on.People is an important calss of object detection.People detection and people counting in video play an important role in security and economic construction planning of smart city.The You Look Only Once(YOLO)is a state-of-the-art real-time object detection algorithm.Compared with other mainstream real-time object detection methods,the YOLO algorithm has faster detection speed and better real-time performance.Although the series of YOLO detection algorithm has good performance,the detection model is very complicated and occupies a large storage space,and the real-time operation requires powerful computing resources support of Graphics Processing Unit(GPU),which can not be used on platforms with limited computing resources and memory.In recent years,the deployment of detection algorithms on front-end embedded or mobile devices has become popular.The computing resources and memory resources of such devices usually have certain limitations,which puts higher requirements on the object detection algorithm.Although the Tiny-YOLO-V3 is designed for constrained environments,the model is still large and the actual execution speed is not fast enough.In this paper,the real-time object detection algorithm YOLO is studied,focusing on the optimization and compression of related structural models.The real-time people detection and counting algorithm and real-time object detection algorithm for constrained environments are studied.These algorithms are improved and optimized based on YOLO and Tiny-YOLO-V3 algorithm.The specific works are organized as follows:(1)A real-time YOLO-based people counting approach named as YOLO-PC is studied.YOLO-PC has a more detailed division of the detection grid cells,which obtains more detection bounding boxes and higher detection confidence.Based on the method of boundary selection,the algorithm counts people in a targeted manner.It can effectively count people at entrances and exits of elevators and buildings,etc.It can count the number of people at a certain time,and can also count the flow of people in a period of time.Experimental results show that the YOLO-PC algorithm based on multi division boundary selection method has real-time detection speed and high precision in the task of people counting.(2)A real-time squeeze YOLO-based people counting approach named as S-YOLO-PC is studied.YOLO-PC has a large number of parameters due to its multi division.So it takes a lot of storage space on the device,and makes the model upload and download on the network timeconsuming.S-YOLO-PC introduces the fire modules of SqueezeeNet to optimize the overall network structure,and optimizes the use of the fire layer modules to achieve the compression of the number of parameters in the model.The experimental results show that S-YOLO-PC reduces the model storage space compared with YOLO-PC and the mAP has almost no loss.(3)A real-time object detection method for constrained environments based on the TinyYOLO-V3 named as Tinier-YOLO is studied.Model parameters and calculations are reduced by introducing a lot of fire modules in Tiny-YOLO-V3.When the model is compressed to a certain extent,it will affect the AP.Tinier-YOLO further uses the dense connections of the fire modules and more fine-grained features to improve the AP of the model,and further removes some parts of BN operations of the fire modules to optimize the overall performance of the network.Finally,the model parameters are greatly reduced in the case of not lossing the detection AP or even improving the detection AP,reducing a large amount of model storage space and calculation amount,and realizing fast and real-time detection performance on the constrained environments.The experimental results show that Tinier-YOLO is much less timeconsuming and space-consuming compared with Tiny-YOLO-V3,and is more suitable for constrained environments.Tinier-YOLO achieved 67.5% Mean Average Precision(mAP)on the VOC 2007 test set,which is better than Tiny-YOLO-V3.On the Microsoft Common Objects in Context(MS COCO)data set,when the Intersection-over-Union(IoU)is 0.5,the mAP is the same as Tiny-YOLO-V3.However,under more stringent evaluation conditions which the IoU is 0.75 or 0.50 to 0.95,the mAP of Tinier-YOLO is better than Tiny-YOLO-V3.In general,Tinier-YOLO is superior to Tiny-YOLO-V3 in the performance of model space occupancy,calculation,detection speed and mAP.Based on the real-time object detection algorithm YOLO,this paper proposes a real-time people detection and counting algorithm YOLO-PC,which has good performance in detection speed and precision.Based on further compression optimization of YOLO-PC,S-YOLO-PC is proposed.The detection algorithm becomes more efficient and reduces a large number of redundant parameters.And YOLO is further optimized based on Tiny-YOLO-V3,a real-time object detection algorithm Tinier-YOLO for constrained environments is proposed.This algorithm not only makes the model smaller,but also has less computation and higher speed.As a smaller and faster object detection algorithm,Tinier-YOLO is more suitable for constrained environments where the mAP of many classes of objects can be greatly improved.