Convolutional Networks Model Optimization Oriented Bus Passenger Detection

Bus passenger detection is an important part of building an intelligent public transportation system.The detection of bus passengers based on image information has become a research hotspot in the field of computer vision.Passenger detection problems in public transport scenes have high requirements on the detection accuracy and detection speed of the algorithm.Although traditional image-based detection methods can meet the speed requirements,their detection accuracy is inadequate.The convolutional neural network has high detection accuracy in the field of object detection,but it has a huge amount of calculation and consumes many resources.Therefore,it is of great practical significance to develop a set of bus passenger detection technology with high speed and accuracy.For this reason,the following research has been done in this paper around the detection accuracy and detection speed.This thesis first analyzes the bus passenger detection task from both the camera angle and the detection object,and determines the data collected vertically from the camera directly above the vehicle door as the detected image of the passenger detection.The shoulder and head area of the passenger in the image are taken as detection object.Then,based on the characteristics of passenger detection in the bus scene,the object detection algorithm of Faster R-CNN,SSD and YOLO is analyzed,and a lightweight tiny yolo model under YOLO detection algorithm is adopted.Finally,in the experiment,the model achieved a detection accuracy of 94% and a detection speed of 1fps.Compared with the traditional method using manual features,the detection accuracy is greatly improved,but the detection speed is slow and cannot be practically applied.In order to improve the detection speed of the convolutional network and maintain the detection accuracy,this thesis next studies the structure and channel of the convolutional network model.This thesis analyzes the time-consuming of each function in the detect program,and it is found that the function with the highest time-consuming function is to implement the convolution-related matrix multiplication function.To solve this problem,a depth-wise separable convolution method is used to divide standard convolution into depth-wise convolution and point-wise convolution.The results show that after replacing the standard convolutions in the network with depth-wise convolution and point-wise convolution,the detection speed has greatly improved,meanwhile there is almost no drop in the detection accuracy.After the optimization of the convolutional network structure,the channel aspects are studied next.Firstly,the convolution layer is determined according to the amount of calculation and parameter,then the contribution of the channel to the detection results is evaluated based on the effect which that one channel is deleted makes on the detection results,and it is determined which channels have priority to be deleted according to the size of the contribution.Finally,the compression algorithm is used to compress the convolution layer.Experimental results show that the method based on the compression process can increase the detection speed and maintain the detection accuracy.After optimization of the network structure and channel compression,the model is 4.7 times faster than the unoptimized tiny yolo model,and only one percentage point lower in detection accuracy,resulting in better optimization results.