Citation Context

...bed, blocks the sensor’s view of the end of the fall, and thus the entire fall is not visible. These occlusions frequently occur at a home environment since a room contains furniture and objects that could be placed between the subject and the camera. Kinect cameras are used to capture the occluded fall benchmark dataset, and the detection is thus based on depth videos. Several sensor-based approaches have been proposed for fall detection, some recent reviews include [5, 6]. An accelerometer [7–9] is the most commonly used sensor, and it is often combined with other devices such as gyroscopes [9] and microphone [8]. These approaches do not suffer from the occlusion problem. However, these methods require subjects to actively cooperate by wearing the sensors, which can be problematic and possibly uncomfortable (e.g., wearing sensors while sleeping, detecting falls during a night trip to the restroom). Several vision-based methods have been proposed for fall detection. They can be broadly classified into two categories: 2D-based and 3D-based. [10–13] use 2D appearance-based features to detect falls. These methods use a single camera and 2 Z. Zhang, C. Conly, and V. Athitsos are viewpoin...

Citation Context

...e. In this work we focus on the task of vision-based occluded fall detection. An occluded fall occurs when an object, such as a bed, blocks the sensor’s view of the end of the fall, and thus the entire fall is not visible. These occlusions frequently occur at a home environment since a room contains furniture and objects that could be placed between the subject and the camera. Kinect cameras are used to capture the occluded fall benchmark dataset, and the detection is thus based on depth videos. Several sensor-based approaches have been proposed for fall detection, some recent reviews include [5, 6]. An accelerometer [7–9] is the most commonly used sensor, and it is often combined with other devices such as gyroscopes [9] and microphone [8]. These approaches do not suffer from the occlusion problem. However, these methods require subjects to actively cooperate by wearing the sensors, which can be problematic and possibly uncomfortable (e.g., wearing sensors while sleeping, detecting falls during a night trip to the restroom). Several vision-based methods have been proposed for fall detection. They can be broadly classified into two categories: 2D-based and 3D-based. [10–13] use 2D appear...

Citation Context

...the restroom). Several vision-based methods have been proposed for fall detection. They can be broadly classified into two categories: 2D-based and 3D-based. [10–13] use 2D appearance-based features to detect falls. These methods use a single camera and 2 Z. Zhang, C. Conly, and V. Athitsos are viewpoint-dependent. Moving a camera to a different viewpoint (especially to a different height from the floor) would require collecting new training data for that specific viewpoint. The 3D features for fall detection can be extracted from a calibrated multi-camera system [14–16] or from depth cameras [1, 3, 4, 2]. Multicamera calibrated systems require time-consuming external camera calibration while depth cameras based systems do not. Most of these methods did not report any information about the robustness of the system to occlusions aside from [16] and [2]. Auvinet et al. [16] used multiple cameras to cover the whole space so that no occlusion occurred for at least one of these cameras. Thus, the total volume information in the scene can be reconstructed based on the multiple-cameras network. This solution, however, is expensive and difficult to set up (multiple calibrated and synchronized cameras ...

Citation Context

...the restroom). Several vision-based methods have been proposed for fall detection. They can be broadly classified into two categories: 2D-based and 3D-based. [10–13] use 2D appearance-based features to detect falls. These methods use a single camera and 2 Z. Zhang, C. Conly, and V. Athitsos are viewpoint-dependent. Moving a camera to a different viewpoint (especially to a different height from the floor) would require collecting new training data for that specific viewpoint. The 3D features for fall detection can be extracted from a calibrated multi-camera system [14–16] or from depth cameras [1, 3, 4, 2]. Multicamera calibrated systems require time-consuming external camera calibration while depth cameras based systems do not. Most of these methods did not report any information about the robustness of the system to occlusions aside from [16] and [2]. Auvinet et al. [16] used multiple cameras to cover the whole space so that no occlusion occurred for at least one of these cameras. Thus, the total volume information in the scene can be reconstructed based on the multiple-cameras network. This solution, however, is expensive and difficult to set up (multiple calibrated and synchronized cameras ...

Citation Context

...nd 2 Z. Zhang, C. Conly, and V. Athitsos are viewpoint-dependent. Moving a camera to a different viewpoint (especially to a different height from the floor) would require collecting new training data for that specific viewpoint. The 3D features for fall detection can be extracted from a calibrated multi-camera system [14–16] or from depth cameras [1, 3, 4, 2]. Multicamera calibrated systems require time-consuming external camera calibration while depth cameras based systems do not. Most of these methods did not report any information about the robustness of the system to occlusions aside from [16] and [2]. Auvinet et al. [16] used multiple cameras to cover the whole space so that no occlusion occurred for at least one of these cameras. Thus, the total volume information in the scene can be reconstructed based on the multiple-cameras network. This solution, however, is expensive and difficult to set up (multiple calibrated and synchronized cameras are needed). Rougier et al. [2] classified each frame into two classes: the person is totally occluded or not. If it is an occluded case, the velocity of the person just before the occlusion occurs is used to determine whether a fall has occur...

Citation Context

...the restroom). Several vision-based methods have been proposed for fall detection. They can be broadly classified into two categories: 2D-based and 3D-based. [10–13] use 2D appearance-based features to detect falls. These methods use a single camera and 2 Z. Zhang, C. Conly, and V. Athitsos are viewpoint-dependent. Moving a camera to a different viewpoint (especially to a different height from the floor) would require collecting new training data for that specific viewpoint. The 3D features for fall detection can be extracted from a calibrated multi-camera system [14–16] or from depth cameras [1, 3, 4, 2]. Multicamera calibrated systems require time-consuming external camera calibration while depth cameras based systems do not. Most of these methods did not report any information about the robustness of the system to occlusions aside from [16] and [2]. Auvinet et al. [16] used multiple cameras to cover the whole space so that no occlusion occurred for at least one of these cameras. Thus, the total volume information in the scene can be reconstructed based on the multiple-cameras network. This solution, however, is expensive and difficult to set up (multiple calibrated and synchronized cameras ...

Citation Context

...the restroom). Several vision-based methods have been proposed for fall detection. They can be broadly classified into two categories: 2D-based and 3D-based. [10–13] use 2D appearance-based features to detect falls. These methods use a single camera and 2 Z. Zhang, C. Conly, and V. Athitsos are viewpoint-dependent. Moving a camera to a different viewpoint (especially to a different height from the floor) would require collecting new training data for that specific viewpoint. The 3D features for fall detection can be extracted from a calibrated multi-camera system [14–16] or from depth cameras [1, 3, 4, 2]. Multicamera calibrated systems require time-consuming external camera calibration while depth cameras based systems do not. Most of these methods did not report any information about the robustness of the system to occlusions aside from [16] and [2]. Auvinet et al. [16] used multiple cameras to cover the whole space so that no occlusion occurred for at least one of these cameras. Thus, the total volume information in the scene can be reconstructed based on the multiple-cameras network. This solution, however, is expensive and difficult to set up (multiple calibrated and synchronized cameras ...

Citation Context

...sor’s view of the end of the fall, and thus the entire fall is not visible. These occlusions frequently occur at a home environment since a room contains furniture and objects that could be placed between the subject and the camera. Kinect cameras are used to capture the occluded fall benchmark dataset, and the detection is thus based on depth videos. Several sensor-based approaches have been proposed for fall detection, some recent reviews include [5, 6]. An accelerometer [7–9] is the most commonly used sensor, and it is often combined with other devices such as gyroscopes [9] and microphone [8]. These approaches do not suffer from the occlusion problem. However, these methods require subjects to actively cooperate by wearing the sensors, which can be problematic and possibly uncomfortable (e.g., wearing sensors while sleeping, detecting falls during a night trip to the restroom). Several vision-based methods have been proposed for fall detection. They can be broadly classified into two categories: 2D-based and 3D-based. [10–13] use 2D appearance-based features to detect falls. These methods use a single camera and 2 Z. Zhang, C. Conly, and V. Athitsos are viewpoint-dependent. Moving...

Citation Context

...e. In this work we focus on the task of vision-based occluded fall detection. An occluded fall occurs when an object, such as a bed, blocks the sensor’s view of the end of the fall, and thus the entire fall is not visible. These occlusions frequently occur at a home environment since a room contains furniture and objects that could be placed between the subject and the camera. Kinect cameras are used to capture the occluded fall benchmark dataset, and the detection is thus based on depth videos. Several sensor-based approaches have been proposed for fall detection, some recent reviews include [5, 6]. An accelerometer [7–9] is the most commonly used sensor, and it is often combined with other devices such as gyroscopes [9] and microphone [8]. These approaches do not suffer from the occlusion problem. However, these methods require subjects to actively cooperate by wearing the sensors, which can be problematic and possibly uncomfortable (e.g., wearing sensors while sleeping, detecting falls during a night trip to the restroom). Several vision-based methods have been proposed for fall detection. They can be broadly classified into two categories: 2D-based and 3D-based. [10–13] use 2D appear...