A Mobile Robot Generating Video Summaries of Seniors' Indoor Activities

TL;DR

This paper presents a system for generating concise video summaries of seniors' indoor activities captured by a mobile robot, using pose estimation, person identification, and action recognition to handle long, redundant, and challenging video data.

Contribution

The novel approach integrates pose estimation, person identification, and action recognition to improve video summarization of indoor activities from a moving robot.

Findings

Effective summarization of seniors' activities in challenging indoor videos

Accurate detection and tracking of seniors using pose estimation and person ID

Generation of diverse, representative keyframes for activity summaries

Abstract

We develop a system which generates summaries from seniors' indoor-activity videos captured by a social robot to help remote family members know their seniors' daily activities at home. Unlike the traditional video summarization datasets, indoor videos captured from a moving robot poses additional challenges, namely, (i) the video sequences are very long (ii) a significant number of video-frames contain no-subject or with subjects at ill-posed locations and scales (iii) most of the well-posed frames contain highly redundant information. To address this problem, we propose to \hl{exploit} pose estimation \hl{for detecting} people in frames\hl{. This guides the robot} to follow the user and capture effective videos. We use person identification to distinguish a target senior from other people. We \hl{also make use of} action recognition to analyze seniors' major activities at different moments, and develop a video summarization method to select diverse and representative keyframes as summaries.

Tables

Table 1. Table 1: Statistics of experimental videos. The subjects’ ages are shown beside their genders.

Video	Subjects	Duration	#Frame (Total/well-posed)
1	male (79)	6h 46m	198689 / 8093
	female (74)
	female (41)
2	female (94)	1h 44m	50634 / 19971
	female (31)
3	female (70)	7h 42m	191183 / 11563

Table 2. Table 2: Execution time for video 2. The execution time for videos 1 and 3 is proportional to their frame numbers.

Method	VSUMM	DPP	DR-DSN	Proposed
Time	11.05s	11m 32s	11m 27s	0.18s

Full text

1 Introduction

\justify

With a large portion of the population becoming aged, we aim to investigate the feasibility of applying video summarization techniques using a social robot to help family members care about seniors living alone. Numerous video summarization methods and interactive robots have been independently developed, but they were never put together. Thus, we investigate their limitations and propose our solution to the widely growing demand.

{marginfigure}

[0pc]

System architecture. Due to the limited computational resources available on a robot, we extend them by a high-performance computer. We transmit data between the robot and the computer via the wireless connection to ensure the robot’s moving capability.

{marginfigure}[0pc]

(a) (b) (c)

(d) (e)

Facial expressions shown on our robot. (a) default_still, if no human is observed. (b) expecting, if the robot is seeing a user’s back. (c) active, if the robot is seeing a user’s eyes. (d)(e) aware_left and aware_right, when the robot is turning left or right to look for a user.

2 Related Work

Our study covers the topics of mobile robots and video summarization because we use a mobile robot to capture videos and analyze the videos to generate summaries.

Autonomous Mobile Robots are capable of navigating an uncontrolled environment and moving around to carry out given tasks. If those robots interact and communicate with humans and their tasks are to improve the quality of users’ life, they belong to social robots. Many studies have shown that social robots are good tools to meet the elderly individual needs and requirements [8]. With the advance of technology, social robots have been extended from zoomorphic robots to humanoid mobile robots equipped with multiple sensors and advanced intelligence [3, 12]. In this paper, we assign a new task of generating video summaries to social robots and develop a solution.

Video Summarization methods analyze videos to create summaries. It is part of machine learning and data mining. The main idea of summarization is to find a diverse and representative subset of the entire input data. Several summary formats are available including keyframes, skimmed videos, time-lapsed videos, and video synopsis. Among them, the keyframes are widely used because they are simple and ease to consume [4, 13].

3 System Architecture

We use a commercial robot Zenbo, as shown in Figure 1, to capture videos. It uses an ultra-low-voltage CPU to reduce power consumption, but the processing power of the CPU is insufficient to analyze videos in real time. Thus, we transmit captured frames via Wi-Fi to a powerful computer to analyze them and return results to the robot.

The robot has a touch screen on its head, which serves as an input UI and also shows facial expressions for human-computer interaction. The robot’s camera is at the upper center boundary of the screen. The robot’s OS is a modified Android system with a set of additional APIs to retrieve sensor data and control the robot’s movements, neck gestures, and facial expressions. There are 24 built-in expressions rendered by OpenGL. We select five of them, as shown in Figure 1, to interact with users. Our software implementation includes three parts, an Android-based app running on the robot to transmit captured frames to a server and receive analyzed results from the server to control the robot’s actions, a C++ program running on the server to analyze images, and an offline Python program to generate summaries by selecting keyframes from all captured frames.

4 Human Detection

Because our summaries focus on human activities, we utilize existing human detection algorithms to find our human targets. Figure 1 shows the relationship among those algorithms. Given an input video frame, we use a real-time pose estimation algorithm OpenPose [2] to find people. We use a pose estimation method rather than a pedestrian or object detection method because we need human body landmark coordinates to control our robot’s camera view angle. We want our robot to capture well-posed frames so we can generate high-quality summaries. In contrast, bounding boxes reported by a pedestrian or object detection method do not analyze human body parts.

Because pose estimation methods do not discern people’s identities, we use a person identification method PSE-ECN [10] to prevent false-positives and determine our target person among people on an image. {marginfigure}[-7pc]

Landmark points.

{marginfigure}[0pc]

Robot-side data flow for movement control.

5 Rule-based Robot Movement Control

OpenPose reports 18 body landmark coordinates as shown in Figure 4, but some may be unavailable if the body parts are invisible. Regarding our target person’s coordinates, if any facial landmark point is available, we adjust the robot’s body orientation and neck pitch angle to change the robot’s camera view angle and move the landmark point(s) towards the central upper part of the upcoming frames to create well-posed ones for summarization. If all facial landmark points are invisible but the neck point is visible, we raise the robot’s head to look for a face. In addition, we also use landmark points to estimate the distance between our robot and the target user to control the robot’s forward movement. We compute the average distances from the neck point to the two hip points because they are the longest connected distances and they distort less than other body parts do in different poses. We move our robot toward the target person until 2 meters apart, which is a proper distance to take images with adequate human size. When our robot cannot detect a person, it will turn left or right 30 degrees depending on the last position of a person visible in past frames. If the robot turns around but still cannot detect a person, we set the robot’s neck pitch angle 15 degree vertically because this degree ensures the robot to detect distant people. If the robot turns around twice but still does not detect a person, we let the robot wait in an idle mode for 15 minutes to save energy until a person appears in front of its camera. Figure 4 shows the overall flow of movement control.

6 Content Filter

Our images are captured from a moving camera in an indoor space so many of them are blurred and improper to be selected as summaries. To remove them, we use the variance-of-Laplacian [9] method and set a threshold. We also use the aforementioned OpenPose and PSE-ECN methods to ignore ill-posed frames, including the ones without people or with people but too small, cropped, at corners, or whose faces are invisible. Figure 7 shows examples for 6 cases.

7 Video Summarization

We use keyframes as our summary format due to its efficiency to consume information at a glance. We propose a method to select keyframes by temporally clustering well-posed frames and then selecting one representative frame out of a cluster in terms of human actions because we expect the summary to not only diversely cover seniors’ daily activities but also show the representative ones. To do it, let $\{v_{i}\}$, $i\in$ (1,$n$), be the well-posed frames and $\{t_{i}\}$ be their timestamps, $t_{i}<t_{i+1}$. Let $k$ be the number of keyframes in our summary. We group $\{v_{i}\}$ into at least $k$ clusters $\{C_{j}\}$, $j\in$ (1,$m$) and $m\geq k$. Initially we let $C_{1}$ contain the first frame $v_{1}$. For any other frames $v_{i}$, $i\in$ (2,$n$), we assign their clusters by the temporal difference from its previous frame, i.e., assume $v_{i-1}\in C_{j}$, we assign v_{i}\in\left\{\begin{array}[]{ll}C_{j}&\mbox{if t_{i}-t_{i-1}<h};\\ C_{j+1}&\mbox{else},\end{array}\right. where $h$ is a temporal gap threshold. In order to produce at least $k$ clusters, We iteratively adjust h=\left\{\begin{array}[]{ll}2h&\mbox{if m\geq k};\\ \frac{h}{2}&\mbox{else,}\end{array}\right. until $m(h)\geq k$ and $m(2h)<k$ where $m(h)$ is the number of clusters determined by $h$. If $m(h)>k$, we disregard the small $m(h)-k$ clusters and only use the $k$ large ones.

We extract frame features as the probabilities of 157 predefined indoor actions generated by a SqueezeNet [6] model pre-trained on the Charades dataset [11], which aims to recognize human actions from a single frame. We compute the mean features of a cluster, and select the frame with the closest distance to the cluster mean, as the cluster’s keyframe. {marginfigure}[0pc]

(a) Blurred

(b) Eyes invisible

(c) People absent

(d) Forehead cropped

(e) People at corners

(f) People too small

Examples of ill-posed frames, which are removed by our content filter.

8 Experiments

We conduct experiments in three families, and their statics are shown in Table 1. We set the keyframe number $k$ as 8 and the initial threshold $h$ as 60 seconds. We compare the proposed video summarization method with three existing ones: VSUMM [4], DPP [5], and DR-DSN [14]. All of them are programmed in Python and run on a machine equipped with a 3.4GHz quadratic core CPU and their execution time is shown in Table 2. We use the publicly available code of OpenPose implemented in OpenVINO [7] but temporally disable PSE-ECN because we have not fully integrated it into our program.

Qualitative comparisons are enclosed in Figure 2. A set of 8 key frames (a-h) selected by the 4 different methods are shown on the right while the histogram of well-posed frames and the time of keyframes selection are shown on the left. The proposed method prevents the problem of consecutive similar keyframes arising in other methods because it exploits timestamps and a threshold $h$ to partition frames into temporally disjoint clusters, which are separated from each other with a gap that tends to isolate different activities. In contrast, VSUMM and DPP ignore temporal information and their keyframes are not sufficiently diverse. DR-DSN generates summaries containing diverse keyframes but neglects representative events of eating a lunch box in video 1, and sitting on a sofa in videos 2 and 3. The proposed method generates diverse and representative summaries, but we have not fully implemented the person identification component, which results in a person on TV shown in video 3’s summary, as the images 12(g)(h) of Figure 2. Our code is available at https://github.com/yangchihyuan/RobotVideoSummary.

9 Conclusion and Future Study

The paper presents an effective method to generate video summaries using a social robot for family members to care about seniors living alone. We use a pose estimation method to detect humans to control the robot’s movements to capture well-posed frames. We use human pose and image quality information to disregard ill-posed frames and develop a summarization method to generate diverse and representative summaries. Experimental results show that our summaries prevent the problems of redundancy and unrepresentative keyframes generated by existing methods.

Our immediate plan is to integrate a person identification algorithm into our system to ensure the robot to keep track on a target user. During our experiment, our users express strong demand for fall detection and immediate notification, which are important features. In addition, we plan to take preferences into account to create personalized summaries.

10 Acknowledgements

This research was supported in part by the Ministry of Science and Technology of Taiwan (MOST 108-2633-E-002-001, 107-2218-E-002-009, 107-2811-E-002-575), National Taiwan University (NTU-108L104039), Intel Corporation, Delta Electronics and Compal Electronics.