Research On Target Tracking And Expression Recognition Algorithm Based On Convolutional Neural Network

Target tracking and expression recognition algorithms based on convolutional neural networks have been widely used in many fields,such as video surveillance,human-computer interaction,and autonomous driving.However,in fields such as UAV and PTZ automatic tracking,the algorithm not only needs to track the target in real time,but also needs to automatically acquire the target and recognize it.In addition,in some tracking applications,such as tracking criminal suspects,ticket evaders and cheaters,it is also necessary to recognize facial expressions of the tracked person to further determine whether the target is a real offender.Therefore,the research on the target tracking algorithm and the facial expression recognition algorithm(FER)of tracking person have very important research significance and application value.Although current single target trackers can achieve better performance in multiple benchmarks,they still cannot automatically detect the target category.However,due to shortcomings such as missed detection and false detection,the multi-target tracker has poor tracking accuracy and robustness.In addition,most of the above trackers do not have the function of automatically acquiring targets.For FER of tracking targets,the recognition effect of facial expression recognition algorithms based on deep learning in practical applications is still not very satisfactory,and there is still much room for improvement in recognition accuracy and parameters of model.Therefore,in order to solve the problems of the above target tracking and FER,this paper mainly carried out the research of target tracking and FER.The main research work completed are as follows:1.An improved target tracking algorithm based on convolutional neural network(1)A novel target tracking algorithm(YOLOv3-DWSiam RPN)is proposed in this paper,which is mainly composed of detection network(YOLOv3),intelligent selection strategy,deeper and wider siamese network(DWSiam RPN)and Kalman filter.The experimental results show that the algorithm proposed can automatically acquire the target of interest,and track and detect target in real time(26 frames per second)in a complex background environment.The mean Accuracy Precision(m AP)of detection is 55.5 in the COCO,and the Expect Average Overlap(EAO)rate of target tracking is 0.30 in the VOT-2017,which has high detection and tracking accuracy and strong robustness.(2)In order to automatically obtain the target of interest,an intelligent selection strategy is proposed in this paper.According to four influencing factors(category,area,detection accuracy and speed),the score of the object is calculated by this strategy,and the object with the highest score is regarded as the object of interest.Parameters of intelligent selection strategy can be set according to different applications,which has high flexibility and scalability.Since this strategy is the key connection between the detection network and the tracker,it provides a reference value for the effective combination of target detection and tracking algorithms.(3)Aiming at the problems of occlusion and disappearance of the target,this paper combines the Kalman filter algorithm to predict the temporarily occluded target.The experimental results show that the Kalman filter algorithm can effectively solve the problems of occlusion and disappearance of the target moving at a constant speed,which has a certain practical application value.2.A facial expression recognition algorithm based on improved residual network(1)In order to further improve the accuracy of facial expression recognition(FER),a facial expression recognition algorithm(SE-SRes Net18)is proposed based on residual network(Res Net)and squeeze and exception network(SENet)in this paper.Firstly,aiming at the problem that there is less training data in the expression recognition database,two methods(random resize crop and random horizontal flip)are used to enhance the training data.Secondly,in order to improve the recognition accuracy,SENet is embeded in the improved residual network;Finally,in order to prevent over fitting,a dropout mechanism is added between the average pooling layer and the fully connected layer.The experimental results show that the recognition accuracy of SE-SRes Net18 on FER2013 and CK+ is 74.14% and 95.25%,respectively.Compared with the current state-of-the-art expression recognition algorithms,it not only improves the accuracy of facial expression recognition,but also has fewer network model parameters,and the performance has been improved.(2)In order to explore the effect of network depth on the accuracy of FER,multiple convolutional layers are added based on SE-SRes Net18 in this paper.Through this method,a26-layer network model(SE-SRes Net26)is proposed.The experimental results show that,compared with the SE-SRes Net18,the number of layers of SE-SRes Net is further increased,and the accuracy of FER on the FFE2013 and CK+(73.59% and 94.34%)is not improved.(3)In view of the effect of different reduction ratios r of SENet,a network model with r=4(SE-SRes Net18?r4)is proposed based on the SE-SRes Net18.Experimental results show that,compared with the SE-SRes Net18,the reduction ratio r is further reduced,and the accuracy of FER on the FFE2013 and CK+ is not improved(74.06% and 94.44%).3.An online person target tracking and expression recognition systemIn order to verify the performance of the YOLOv3-DWSiam RPN and SE-SRes Net18 algorithms in practical applications,this paper combines the latest Center Fcae face detector to design a novel online person target tracking and expression recognition system.The system consists of target detection and tracking module(YOLOv3-DWSiam RPN),face detection module(Center Face)and expression recognition module(SE-SRes Net18).The experimental results show that the system can automatically acquire an interesting person as target in a complex background environment,and automatically track and recognize facial expressions,which has good potential application value.