Unsupervised spatio-temporal multi-human detection and recognition in complex scene

An algorithm for multi-human detection and recognition in complex scene is proposed. It executes multi-human detection and recognition from spatial domain and time domain. In spatial domain, it establishes the Gaussian mixture background model to obtain target windows by the background difference and the foreground connection judgment. Based on target detection, a new target size estimation method is carried out through the computation of depth of field in the scene. Each target eigenvector is extracted from its contour to input into support vector machine (SVM) to judge the target was human or not. In time domain, a new three-layer bidirectional min-distance data association is proposed. It finds out the forerunner and successor associations of target data in some image sequences based on target position, target size and target gray. It gives each target chain human or unhuman property to assist in completing human recognition. Finally, a spatio-temporal union mechanism is proposed. It presents good result on multi-human detection and recognition.