Classification and Regression Models to Evaluate Human Motion Patterns from Image Sequences
|關鍵字:||人類運動分析;步態;活動;瞌睡;視覺;線性鑑別式;分類;迴歸;human motion analysis;gait;activity;drowsiness;vision-based;linear discriminant analysis;classification;regression|
|摘要:||人類運動分析已被視為在醫學、運動及監視系統中重要的輔助工具。一般來說主要有以感測器為基礎與以視覺為基礎兩種方式來分析人類的運動。感測器為基礎的方式可能造成不連續的紀錄、影響運動形態、不能完整觀察全身及讓受測者不舒服的缺點。早期以視覺為基礎的方法，如運動捕捉系統克服了以上缺失，但仍相當昂貴、耗時並受限於場地。後期的研究雖進一步改善早期的問題，仍有影像分割和參數估測條件嚴苛及特徵不夠有效率的問題，使辨識正確率仍不高。除此，人類運動的迴歸是一重要主題，卻少有研究探討。在本文中，我們發展新的人類運動形態評估系統來處理分類與回歸兩種情形。我們使用非模型的策略來避免影像分割和參數估測的問題，更進一步，我們用線性鑑別式來增加特徵的效率。因為它能分別將組內變異和組間變異予以最小和最大化。藉著分類的技巧，我們設計了空間和時間資訊上的兩個新指標。最後再以線性回歸來完成人類運動的回歸分析。本系統能對人類步態形態做分類及異常程度預測、辨識人類活動及偵測人類瞌睡狀態。我們的演算法對正常人步態做身份辨識有93.63%的正確率，在辨識正常人和運動失常患者時正確率達到95.49%，而在異常程度預測上我們的方法比對UPDRS Part III分數時的相關係數亦高達0.9109 (p = 0.0000)。對於人類活動辨識，我們採用LDA來辨識外型，並以模糊推論學習活動時的時間特徵，改進傳統方法不足之處。我們的方法經測試有91.78%的正確率，高於最接近鄰域法及HMM法。對於人類瞌睡狀態偵測，我們設計了非侵入式方法來避免傳統方法造成的不便。另外我們用模糊積分結合PERCLOS和LDBF兩項重要瞌睡因子。本系統經驗證有高達95.1%的正確性。我們的方法也實現於以PC為基礎的即時系統上，並對駕駛者精神狀態加以監視並在危險時提出警告。總結來說，我們開發的系統解決了四項在人類動作分析上的重要主題：包含步態辨識、步態異常程度預測、人類活動辨識及人類瞌睡狀態偵測。在本文中，我們以視覺的方法，並只採用一部攝影機，因此，我們的方法乃是非侵入式兼顧了舒適性、高正確率、低成本及易於實現。由於本論文的方法均經實驗驗證，因此，我們的系統是一個客觀上有效且適用於輔助人類的運動形態分析。
Human motion analysis has been treated as an investigative and diagnostic assistant tool in medicine, sports, and surveillance. Generally speaking, there are two categories of approaches for human motion analysis: sensor-based and vision-based. The drawbacks of sensor-based methods are possible discontinuous records, affection of motion patterns, incomplete observation of the motion of the body, and uncomfortableness. Early vision-based methods like motion capturing overcome above disadvantages but still are expensive, time-consuming for preparation, limited to specific lab. Recent vision-based methods improved above demerits but might be critical to segmentation and parameter estimation, use inefficient features, and thus be low accuracy. In addition, regressions in human motion analysis is essential but there are very limited studies on it. Therefore, we are intended to develop new evaluation systems for the analysis of human movement patterns for both classification and regression. We use a model-free strategy and thus avoid the critical demands of segmentation and parameter estimation. Furthermore, we use linear discriminant analysis to increase the feature efficiency by maximizing and minimizing the between- and within-group variations. Regression is also achieved by assessing spatial and temporal information through classification and finally by using these two new indices for linear regression. Our systems are designed to be capable of classifying human gait patterns, predicting the abnormality degree a given gait image sequences, recognizing human activities, and detecting the drowsiness of people. For judging the identification of normal people, the accuracy was 93.63%. For separating PD gaits from normal people, it was as good as 95.49%. According to the experiments, the outcome had correlation to the sum of the UPDRS Part III sub-scores with r = 0.9109. For human activity recognition, we improved the disadvantages of traditional approaches by using both shape-based features by LDA classification and temporal correlations by fuzzy inference. The approach was tested to be robust and had a very good recognition accuracy of 91.78%, which was higher than nearest neighbor and HMM. For drowsiness detection, we designed a non-intrusive method to avoid the inconvenience caused by traditional approaches. On the other hand, we also improve the detection accuracy of drowsiness by a fuzzy integral based fusion method combining PERCLOS and LDBF of the eyes. The approach had very high drowsiness detection accuracy of 95.1%. Our method was also implemented to be a PC-based real-time system. The proposed algorithm was implemented in a real-world driver drowsiness detection and warning system. In summary, we developed several models to address four important issues in human motion analysis: the classification and regression of human gaits, the classification of human activity, and the detection of human drowsiness. In this dissertation, we take the vision-based approaches by using only a camera. Thus, our methods are all non-intrusive, easy to implemented, cost-effective, comfortable for subjects. We also verified our models by comparing the performance with traditional methods. As a result, our method was an objective and cost-efficient assistant way to provide accurate human motion analysis. Keywords: human motion analysis, gait, activity, drowsiness, vision-based, linear discriminant analysis, classification, regression.