在行人辨識中,衣著的辨識是一個重要的關鍵點,目前也有非常多相關的研究或應用,提供了不少的技術與理論可以參考,但是大部分的實驗大多使用雜誌或網站所提供的影像,受到環境、陰影與姿勢的影響較少,本文利用現實生活的街景影像進行行人上的衣著屬性辨識,衣著屬性包含衣服顏色、衣服長短袖、褲裙顏色、褲裙分類,並且使用影像中,受到較少遮蔽、整體較完整的行人,以達到較好的辨識效果。 本論文使用較接近現實情形的街景影像,利用人工的方式抓取行人的Bounding Box影像,並且對其進行屬性的標記,利用此影像進行深度學習的訓練,使此模型能夠辨識行人的衣著的相關屬性,並且比較不同的輸入影像與模型架構的調整的效果。
In pedestrians recognition, clothing recognition is an important key point. At present, there are many related researches or application, and many techniques and theories are available for reference. However, most of the experiments use images provided by magazines or websites. It is less affected by the environment, shadows and postures. In this paper uses real-life streetscape images to identify the clothing attributes of pedestrians. The clothing attributes include clothing color, clothing sleeves, pant and skirt color, pant and skirt classification. And use less obscured, overall more complete pedestrians of image, in order to achieve better identification. This paper uses the street view image that is closer to the reality. Therefore, the artificial way to capture the Bounding Box image of the pedestrian, and marking the clothes attributes of pedestrians. The image is used for deep learning training. the model can recognize the clothing of the pedestrian attributes. Finally we compare the effects of different input images and model architecture adjustments.