Study On The Target Detection And Recognition Based On The Analysis Of 3D Trajectory In Complex Scenes

Research of object detection and recognition in complex scenes is an important part in the study of intelligent monitoring system. At present, there are still some problems to be solved in the moving object detection and recognition technology, such as sensitivity to light, object occlusion, low detection rate on congestion condition etc.In this paper, an algorithm of object detection and recognition based on 3D trajectory analysis is adopted. The basic idea of the algorithm is detection-tracking-recognition again. First of all, according to the principle of binocular stereo vision, we calibrated the binocular camera, acquired the inside and outside parameters of camera and got the depth image. Then according to the relation between the target height and depth in the depth image, we used the method of searching the local maximum depth to extract head area of target and achieved the initial detection of suspected objects. The Kalman filter was used to forecast the area of object matching. We tracked the object by the multi-feature weighted matching criteria and attained the two-dimensional trajectory in the image. Finally, through the three-dimensional camera calibration, we converted the two-dimensional trajectory of suspected target in the image to three-dimensional trajectory in the world coordinate system and used three kinds of 3D trajectory analysis methods to recognize the object, respectively, algorithm based on the detection plane, algorithm based on the Naive Bayesian classification and algorithm based on AdaBoost classification.In this paper, we have tested the three methods of 3D trajectory analysis and object recognition on the real scene. The results show that all of them can meet the real-time requirement and in the presence of light, as well as under the influence of occlusion and crowded conditions, the precision of algorithm based on the detection plane is 91.4%, the precision of algorithm based on Naive Bayesian classification is 94.84%. The algorithm based on AdaBoost classification works best. Its precision can reach 96.93%.