A Real-Time Human Action Recognition System Using Depth and Inertial Sensor Fusion

RGB-D datasets using microsoft kinect or similar sensors: a survey

“…The label information (the forth column) is valuable because it aids in the process of annotation. The data modalities (the fifth column) include color, depth, skeleton and accelerometer, which are helpful for researchers to quickly identify the datasets especially when they work on multi-modal fusion [15,16,58]. Accelerometer data is able to indicate the potential impact of the object and starts an analysis of depth information, at the same time, it simplifies complexity of the motion feature and increases its reliability.…”

“…The collection procedure for one object takes roughly 5 mins. In this procedure, four adjustable lights are put in 16 Sampled RGB (top row) and depth images (bottom row) from Kinect FaceDB [64]. Left to right: neutral face with normal illumination, smiling, mouth open, strong illumination, occlusion by sunglasses, occlusion by hand, occlusion by paper, turn face right and turn face left different places, illuminating the recording environment.…”

Naive Bayesian Fusion for Action Recognition from Kinect

“…The comparison of the existing work with our approach using the UTD-MHAD dataset for the half-subject experiment is illustrated in Table 3. The second experiments are the subject-specific settings in [19]. As each subject performs an action four times, the first two repetitions are used for training and the two remaining repetitions for testing.…”

“…To get action prediction outputs from the feature variables, we train three classifiers: Collaborative Representation Classifier (CRC) [10,14], Sparse Representation Classifier (SRC) [15,16] and Kernel based Extreme Learning Machine (KELM) [17]. These techniques are among the most widely used methods in the literature [9,10,14,15,18,19], as they have shown good performances for activity recognition systems, but as far we know, this is the first time that these three classifiers are fused together to classify action. Finally, we consider a Naive-Bayes approach to combine the classification scores, that shows an improvement in the accuracy of human action recognition when tested on publicly available datasets [14] [20].…”

Research of Action Recognition Methods Based on RGB+D Videos

“…However, the data collected by the Kinect camera not only include RGB data, but also include 3D skeletal data, depth map sequences and infrared videos. Chen et al [4] used a depth camera and an inertial sensor to built a real time action recognition system by a decision-level fusion. Wang et al [5] proposed a algorithm to mine a set of key-pose-motifs to recognize actions from skeletal data by matching a sequence to the motifs of each class and selecting the class that maximizes the matching score.…”