State Estimation And Traiectory Prediction Of Fast Flying Object

Trajectory prediction, interception and hitting back of moving objects have significant research and promotion value in many fields such as sport, military, and industry etc. The technologies involved such as target perception, tracking, dynamic motion modeling, decision-making, motion planning and control are also the key research issues in robot community. In this thesis, we will use the ping-pong robots "Wu" and "Kong" developed by Zhejiang University as the platform to study algorithms for fast vision perception, motion modeling, trajectory prediction, and spin estimation of ping-pong ball. Based on existing researches, the main contributions of this thesis are listed as follows:1. To overcome the difficulties in object recognition of high speed vision system, such as limited pixels occupied by the object, non-uniform distribution of color, disturbance in image due to uneven illumination and pose conditions, a statistic based target recognition method and an encoding based fast contour searching method are proposed. By analying the histogram feature of the target object and the background, the proposed method is capable to deal with the disturbance of uneven illumination, reflection, shadow and imaging blur. By defining the coding sequence of pixels’ position and its edge, and summarizing the regular pattern of successive point and successive edge when following a closed contour, the proposed method effectively accelerates the target searching process. Combining with preprocess techniques such as background subtraction, dynamic search window, down-sampling, and with the advantages of the proposed multi-thread processing framework, the proposed stereo vision system can recognize and locate the target precisely, quickly, effectively and robustly. The experimental results show the ratio of successful recognition for flying ball is above 99%, the contour searching process is 4 times speeded up, the target recognition process cost 0.15ms in average and the 3D location precision is less than 1cm.2. Based on the requirements for large look-ahead precise trajectory prediction, an overall scheme for adaptive state estimation and trajectory prediction of a flying object is proposed. Different from existing modeling method that using constant parameters and discrete form, the proposed method establishes two equivalent forms of the dynamic model:the discrete form for state estimation and the continuous form for parameters identification and trajectory prediction. The two forms share the same parameters relating to the object’s flying state, whose values are trained offline. Thus the vision system can adapt its model with the optimal parameters accordingly online and get precise trajectory prediction results. The experimental results show that when ball flying over 50cm (time before the robot start to move), the prediction error of the bouncing position is less than 1.2cm and the time error is about 4ms.3. To diminish the prediction error in object’s trajectory due to limited information on the object’s spinning state, a novel vision system that can provide both the position and the spin information of a flying ball in a real-time mode with high accuracy is proposed. Using a frame difference based recognition method, the natural brand of the ball can be recognized under normal illumination conditions. Then the 3D pose of the ball can be restored in ball coordinates. Based on the observation and analysis that the rotation axis and angular speed of the spinning ball remain constant during flying, the spin state can be estimated using a weighted-RANSAC fitting method and mean filtering method. Combining both the position and spin information using a force-based dynamic model, accurate trajectory prediction of a spinning flying ball can be achieved via an EKF filter. The experimental results show the average error of spin estimation is less than 0.07 rev/s during 0-60 rev/s, and the proposed system successfully support the robot hitting back spinning balls.Based on the methods introduced in this thesis, an effective and precise vision system is realized. And it successfully servo our humanoid robots’Wu’ and ’Kong’ playing table tennis with human and with each other, the best record is 145 and 114 turns respectively. A brief summary and prospect forecast is presented at the end of this thesis.