Study On Brain Storm Optimization Algorithm And Video-based Remote Workout Ouantification Approach

As important medical information and diagnosis evidences, the processing and analysis of medical images is one of the most important topics in the biomedical field. A lot of optimization problems are included in medical image processing, such as image segmentation, registration, fusion, compression, reconstruction, and so on. Therefore, superior optimization algorithms are crucial in the field of medical image processing.Swarm intelligence based algorithms are stochastic mechanism based optimization algorithms, which can flexibly deal with the swarm internal changes and the searching environment changes, therefore, they are superior in dealing with optimization problem. Recently, swarm intelligence based algorithms, such as particle swarm optimization (PSO) algorithms, have been widely researched and applied into many aspects of medical image processing. Although powerful, classic swarm intelligence based algorithms are all inspired by social behaviours of low natural biological communities, and still suffer from problems such as slow convergence, premature convergence, etc. Since human beings are the most intelligent social animals, it’s natural to expect superior performance in algorithms inspired by human beings’intelligent social activity. However, there are rarely any algorithms inspired by their intelligent social activity, except the recently proposed brain storm optimization algorithm (BSO).In this dissertation, we proposed a novel discussion mechanism based brain storm algorithm (DMBSO) which got inspiration from human beings’creative problem solving process-brain storm process. We further improved the discussion mechanism and introduced the differential step strategy, which enhanced the adaptability of the algorithm and increased the convergence rate. Experiments were implemented on 24 classic benchmark functions, including unimodal, multimodal and composition functions. The new discussion mechanism based brain storm optimization algorithm was comprehensively assessed and compared with BSO and classic swarm intelligence based algorithms, including PSO and differential evolution (DE) algorithms. The simulation results indicated that our new approach outperformed them greatly both in minima finding and the convergence rate.With the development of mobile devices, besides the conventional biomedical fields, such as medical instruments, medical image processing techniques and so on, mobile health has also become one of the hottest topics in this field. Another research topic in this dissertation is to study on video based remote workout quantification methods, and to further apply the BSO algorithm to workout analysis.Objective quantification of workouts has been a hot topic in the field of mobile health. On one hand, it can objectively track a person’s workout process, motivate him/her to adhere to regular workout routine, and change the inactive life style. On the other hand, it can also provide research scientists and health experts with scientific approach. For now, objective quantification of workouts is based on certain professional laboratory equipment or wearable devices, which are not only expensive and inconvenient for the users, but also may result in inaccuracy by the incorrect wearing. In this dissertation, we proposed a novel video-based remote workout quantification approach, which included not only workout recognition, but also automatic repetitions counting and estimation of the workout intensity. This approach was based on optical flow field, which could be used to estimate velocity field of workout after calibration. Velocity field contains plenty of information of workouts. We used popular indoor workouts, sit up, push up, jumping jack and squat as examples in the primary study. More specifically, workout recognition used Hidden Markov Model (HMM) as the classifier and oriented histogram of optical flow field as the features. Experiments were implemented on a video set downloaded from Youtube. The results showed that the accuracy of recognition was high, around 92.5%. For workout repetition counting, the concept of main optical flow was proposed, which could record the velocity change of the entire body during exercise. We detected the local minima of main optical flow to count the repetitions of workouts automatically. Experiments indicated the effectiveness of this method. Moreover, workouts intensity estimation was based on the mechanical energy generated during the process. In this study, method of calculating mechanical energy from optical flow field was proposed which could further reflect the workout intensity. The results calculated from our methods were compared with real data measured by Oxycon Mobile system. The comparison indicated that our approach could reflect the intensity level accurately.In the last part of this dissertation, the new brain storm optimization algorithm was applied into the recognition of workouts. HMM is well known for its superior ability to modal time-series data, and the training part is the key to the accurate recognition. Training is actually to adjust the coefficients to maximize the probability of the observation series observed. Therefore, in essence, HMM’s training is an optimization problem. Traditional Balm-Welch method is widely used to solve this problem, but it may easily be trapped into the local minima, and thus results in recognition failure. In this dissertation, we introduced BSO algorithm to solve the problem of HMM’s training. Results indicated that it effectively improved the training of HMM and as a result improved the recognition accuracy.