Design And Implementation Of Automatic Detection And Tracking System Based On Image Recognition

With the rapid development of information technology and also the computing power of hardware devices,people can identify and extract data information from images through image processing,mathematical modeling and pattern recognition.Currently,target tracking systems are increasingly used in industrial and commercial applications.However,most of these target tracking systems are still limited to the detection of moving targets,the recognition accuracy and real-time performance are not high enough,and only a short-term target tracking can be achieved,and is especially difficult in coping with complex scenes.The target is easier to be lost when it is deformed or occluded.In order to cope with these problems,this paper researched the key algorithms on target detection and tracking,and designed and implemented a target detection and tracking system,which realized a good real-time and robust ability.Firstly,based on the classical YOLO algorithm,we proposed a P-YOLO algorithm to improve the performance in pedestrian recognition.This algorithm adjusted the network structure of YOLO algorithm,increased the density of lateral detection.The target identification and the classification problem are unified to the regression problem.Based on the end-to-end one-time detection,the candidate region selection,feature extraction and candidate information verification are completed in one time through one network training.We performed this algorithm on the 4010 picture of VOC2007 and VOC2012,the recognition accuracy rate was 82.9%,which was 8.6%higher than the YOLO algorithm.Secondly,based on the KCF algorithm,this paper proposed the CL-KCF algorithm.This algorithm improved the process of target tracking into two steps:position prediction and scale prediction.In the step of location prediction,by combining KCF with the tracking model based on color histogram features,an adaptive joint scheme is proposed to improve the tracking accuracy in the scene where the target deformation occurs.In the step of scale prediction,by introducing a one-dimensional scale filter,it is possible to sample different sizes around the position prediction area,thereby selecting the scale with the largest response value as the prediction scale.On this basis,the update strategy of the model is further modified.Furthermore,the reliability of the algorithm is improved by adding the pre-order reliable frame instead of only the latest predicted frame as the training sample.Finally,a confidence test scheme is proposed in the prediction step,to deal with the model drifts which may reduce the predictive accuracy.This scheme can re-determine the target position by the target recognition algorithm and initialize the CL-KCF tracking model,thus improving The effectiveness of algorithm tracking.Finally,based on the above algorithm,this paper designed and implemented a target detection and tracking system.This system is consisted of video analysis module,target detection and tracking module,and camera follow-up module,etc.It can parse the real-time monitoring video from the camera into the standard input matrix data.On the basis of the target detection and tracking,the camera follower module can keep the tracking object always in the central of the video picture.The system experiments showed that the system can effectively detect pedestrians and track them for a long time,which proves the effectiveness and feasibility of the system.