Research On Key Algorithm Of Object Detection And Tracking Based On Visual

Object detection and tracking has always been hot and cutting edge topics in the field ofcomputer vision. It also plays a very important role in the areas of robot navigation, intelligentmonitoring, video compression and medical images. Object detection and tracking are twoclosely related processes, the result of the two can be mutually supporting. For decades, objectdetection and tracking technology has made great strides through the tireless efforts of themajority of scholars. However, due to the complexity of the application environment ofdetection and tracking system (such as lighting, occlusion and other factors), and the diversityof the object itself (such as changes in appearance and shape) has brought serious difficultiesand new challenges for object detection and tracking technology. In order to resolve theseproblems, this paper studied the following two aspects and achieved some results.(1) In order to find the features which are robust to illumination and deformations, wepropose the ways to construct the weak classifier of MB-LBP features and HOG featuresbased on the AdaBoost algorithm. We analyze and compare the HAAR features, MB-LBPfeatures and HOG features on AdaBoost algorithm classification. The above features in theexperiments of landmark detection and recognition systems and pedestrian detectiondemonstrated that hog feature combine with the AdaBoost algorithm can obtain a higherdetection rate and a lower false positive rate. On the basis of the detection result of thecomparison of the landmarks, the paper proposes a combination of AdaBoost algorithm andSVM algorithm landmark recognition method to obtain a higher recognition rate and fasterrecognition speed.(2) This paper draws on the thinking of Tracking-Learning-Detection (TLD) trackingalgorithm and proposed the pedestrian tracking algorithm which is the combination ofdetection based on the offline learning and tracking based on the online learning. The systemdivides the detection module of TLD into two parts: one part is the pedestrian detector ofoffline training and the other part is the pedestrian validator of online real-time updates. Theoffline trained detector locates all pedestrians in the video and the online trained validatordetermines the tracked pedestrian from the detected pedestrians through the previous state.The combination of offline and online learning, not only can guarantee real-time, but also toovercome the difficulties in pedestrian tracking such as the objectives are similar to thebackground, block, the interaction between people (close, integration, swaps) and maketracking more robust.