Spatio-Temporal Scale Coded Bag-of-Words

The Bag-of-Words (BoW) framework has been widely used in action recognition tasks due to its compact and efficient feature representation. Various modifications have been made to this framework to increase its classification power. This often results in an increased complexity and reduced efficiency. Inspired by the success of image-based scale coded BoW representations, we propose a spatio-temporal scale coded BoW (SC-BoW) for video-based recognition. This involves encoding extracted multi-scale information into BoW representations by partitioning spatio-temporal features into sub-groups based on the spatial scale from which they were extracted. We evaluate SC-BoW in two experimental setups. We first present a general pipeline to perform real-time action recognition with SC-BoW. Secondly, we apply SC-BoW onto the popular Dense Trajectory feature set. Results showed SC-BoW representations to successfully improve performance by 2–7% with low added computational cost. Notably, SC-BoW on Dense Trajectories outperformed more complex deep learning approaches. Thus, scale coding is a low-cost and low-level encoding scheme that increases classification power of the standard BoW without compromising efficiency.

Multiple-Object-Tracking Algorithm Based on Dense Trajectory Voting in Aerial Videos

In recent years, UAV technology has developed rapidly. Due to the mobility, low cost, and variable monitoring altitude of UAVs, multiple-object detection and tracking in aerial videos has become a research hotspot in the field of computer vision. However, due to camera motion, small target size, target adhesion, and unpredictable target motion, it is still difficult to detect and track targets of interest in aerial videos, especially in the case of a low frame rate where the target position changes too much. In this paper, we propose a multiple-object-tracking algorithm based on dense-trajectory voting in aerial videos. The method models the multiple-target-tracking problem as a voting problem of the dense-optical-flow trajectory to the target ID, which can be applied to aerial-surveillance scenes and is robust to low-frame-rate videos. More specifically, we first built an aerial video dataset for vehicle targets, including a training dataset and a diverse test dataset. Based on this, we trained the neural network model by using a deep-learning method to detect vehicles in aerial videos. Thereafter, we calculated the dense optical flow in adjacent frames, and generated effective dense-optical-flow trajectories in each detection bounding box at the current time. When target IDs of optical-flow trajectories are known, the voting results of the optical-flow trajectories in each detection bounding box are counted. Finally, similarity between detection objects in adjacent frames was measured based on the voting results, and tracking results were obtained by data association. In order to evaluate the performance of this algorithm, we conducted experiments on self-built test datasets. A large number of experimental results showed that the proposed algorithm could obtain good target-tracking results in various complex scenarios, and performance was still robust at a low frame rate by changing the video frame rate. In addition, we carried out qualitative and quantitative comparison experiments between the algorithm and three state-of-the-art tracking algorithms, which further proved that this algorithm could not only obtain good tracking results in aerial videos with a normal frame rate, but also had excellent performance under low-frame-rate conditions.

Human Action Recognition Based on Foreground Trajectory and Motion Difference Descriptors

Aimed at the problems of high redundancy of trajectory and susceptibility to background interference in traditional dense trajectory behavior recognition methods, a human action recognition method based on foreground trajectory and motion difference descriptors is proposed. First, the motion magnitude of each frame is estimated by optical flow, and the foreground region is determined according to each motion magnitude of the pixels; the trajectories are only extracted from behavior-related foreground regions. Second, in order to better describe the relative temporal information between different actions, a motion difference descriptor is introduced to describe the foreground trajectory, and the direction histogram of the motion difference is constructed by calculating the direction information of the motion difference per unit time of the trajectory point. Finally, a Fisher vector (FV) is used to encode histogram features to obtain video-level action features, and a support vector machine (SVM) is utilized to classify the action category. Experimental results show that this method can better extract the action-related trajectory, and it can improve the recognition accuracy by 7% compared to the traditional dense trajectory method.