Learning cast shadow appearance for human posture recognition

摘要: A shadow-based multi-view system is proposed for human posture recognition.Challenging images are generated due to cast shadow projected on the floor and walls.Handcrafted features failed describe such challenging images.Synthetic data used CNN training avoid lack of enough real data.The allows transfer learning gives high correct classification rate. This paper presents a recognition using camera two infrared light sources. It uses as input combination body silhouette its (invisible eye) shadows. Conventional video-surveillance methods based single can fail infer since different postures look similar under perspective projection. Fortunately, shadows, by lights, offer additional information that cannot be directly captured camera. Each surfaces (e.g. floor, walls, furniture) generating complex projections represent various shapes within same class. These very difficult with traditional handcrafted need somewhat invariant these within-class changes. However, deep convolution neural network (CNN) able learn better representation from large-scale dataset. In absence big dataset, we propose use synthetic classifier. Learning task gap between feature distributions. Thus, normalization technique bridge help classifier generalize data. We evaluated new dataset in our laboratory simulated computer graphics tools. Experimental results validated efficiency model against other conventional methods. Furthermore, shadows had performance than only expected.