People Orientation Recognition by Mixtures of Wrapped Distributions on Random Trees

Abstract

The recognition of people orientation in single images is still an open issue in several real cases, when the image resolution is poor, body parts cannot be distinguished and localized or motion cannot be exploited. However, the estimation of a person orientation, even an approximated one, could be very useful to improve people tracking and re-identification systems, or to provide a coarse alignment of body models on the input images. In these situations, holistic features seem to be more effective and faster than model based 3D reconstructions. In this paper we propose to describe the people appearance with multi-level HoG feature sets and to classify their orientation using an array of Extremely Randomized Trees classifiers trained on quantized directions. The outputs of the classifiers are then integrated into a global continuous probability density function using a Mixture of Approximated Wrapped Gaussian distributions. Experiments on the TUD Multiview Pedestrians, the Sarc3D, and the 3DPeS datasets confirm the efficacy of the method and the improvement with respect to state of the art approaches

Upper Body Pose Estimation for Team Sports Videos Using a Poselet-Regressor of Spine Pose and Body Orientation Classifiers Conditioned by the Spine Angle Prior

“…ing player (orange arrow in Fig. 3 (a)) as in previous work on surveillance [11], [12], but also estimates the 2D spine pose of the players (purple line in Fig. 3 (a)), which consists of the head center location and the pelvis center location even during the bending poses.…”

“…Specifically, sports players in team sports videos have more body tilt variations than pedestrians poses because they tend to bend their (upper) body while in defensive actions or some specific actions (e.g., passing action in soccer). Larger body tilt variations make the body orientation problem more difficult because the previous body orientation estimators during pedestrian tracking or detection [11], [12] depend on the alignment of the input window by the pedestrian detector for only standing walking poses.…”

A Two-Stage Approach for Bag Detection in Pedestrian Images

“…There are some previous works [10,11,[17][18][19] on other pedestrian attributes recognition in a person image for "is-male", "has-hat", "has-shorts", "has-vnecks", "pedestrian orientation", etc. The approaches for pedestrian attributes recognition can be broadly categorized in two directions.…”

“…The approaches for pedestrian attributes recognition can be broadly categorized in two directions. In one direction, some works train discriminative models using Support Vector Machine (SVM) [11], Adaboost [19] and Random Forest [18] to recognize pedestrian attributes with a feature vector extracted from full body images. However, they neglect that bag regions exist in bag images.…”

Multiview feature distributions for object detection and continuous pose estimation

“…They typically involve voting in the 3D pose space followed by averaging [21] or probabilistic smoothing schemes [1,25,29], leading to a precision beyond the resolution of viewpoints given as training examples. While those simple techniques have sometimes given very interesting results, we rather chose, in the work presented here, to explicitly detect, and include in the model, the changes of appearance between the discrete viewpoints seen during training (practically, how image features translate in the image, and thus how the appearance "deforms" between neighbouring viewpoints).…”