Spatio-activity based object detection

Spatio-activity based object detection

Abstract

We present the SAMMI lightweight object detection method which has a high level of accuracy and robustness, and which is able to operate in an environment with a large number of cameras. Background modeling is based on DCT coefficients provided by cameras. Foreground detection uses similarity in temporal characteristics of adjacent blocks of pixels, which is a computationally inexpensive way to make use of object coherence. Scene model updating uses the approximated median method for improved performance. Evaluation at pixel level and application level shows that SAMMI object detection performs better and faster than the conventional Mixture of Gaussians method.

1Introduction

The field of object detection has matured to the extent that its results are sufficiently robust to be deployed in the real world. Systems using object detection assist in mitigating security threats and gathering marketing information. Research focus has shifted to further analysis making use of object detection results, such as behavior analysis. However, with the large number of surveillance cameras installed each year, scalability of analysis solutions becomes a problem when one server is connected to many cameras. Performing real-time video analysis is not possible without investing significantly in extra infrastructure, such as additional servers or dedicated analysis boxes.

As much of the resource usage in a video analysis is spent on the underlying object detection, it makes sense to revisit detection techniques and make them more suitable for deployment in large operations. Object detection, and especially foreground/background separation, is the bottle neck in an analysis system because the foreground separation needed for object detection is performed on every pixel in the frame. After objects have been found, video analysis can focus at the higher level of objects. Obviously, there are many less objects in the average frame than there are pixels. The problem could be addressed by reducing the sampling. For example, analysis could be done at a lower frame rate than the camera’s capture rate, or processing could be performed on a selected region of interest only. However, such approaches are just shifting the problem and do not make optimal use of the information provided by the cameras.

In this paper, we present a lightweight object detection technique that still has a high level of accuracy and robustness. In the context of this paper, we assume that the camera has a fixed field of view. Yet we do expect that an object detection method reinitialize quickly when the field of view changes in the case of Pan-Tilt-Zoom (PTZ) cameras. Also, we consider objects in the scene that are non-transient during an entire recording, e.g. a parked car, to be part of the background. Transient objects are considered foreground. A foreground object may be stationary for part of the recording, while the background may contain movement, e.g. a swaying tree.

The paper is organized as follows. In section ?, previous work in the field is described. In section ?, a general overview of the system and context in which the spatio-activity based object detection operates is given. In section ?, we present the details of our SAMMI (Spatio-Activity Multi-Mode with Iterations) method. Finally, in sections ? and ?, we evaluate the method and draw conclusions.

2Previous work

The general concept of background modeling that is the basis for our approach is well known in literature. The basic assumption is that background is static and doesn’t change, and that anything new must be foreground. Early approaches, e.g. [4], update a reference frame to be robust to changes in the scene or the appearance of the scene, e.g. lighting changes, and do background subtraction with incoming frames. Background modeling was popularized by Stauffer and Grimson’s Mixture of Gaussians approach [6]. The Mixture of Gaussians approach does not maintain one model per pixel, but several models in the form of a Gaussian. The models are not statically classified as foreground or background, but are dynamically classified after updating the model with the information from a new frame. More than one model can be background, allowing for modeling of multi-modal backgrounds, such as swaying trees.

Li and Sezan [1] have used information about the neighbors of pixels being classified in order to address the problem of spatial and temporal aliasing on cameras. They assume that the probability that a pixel is background depends on the classification of neighboring pixels. The disadvantage of this approach is that it results in fragmentation and erosion of foreground objects. For this reason, Li and Sezan apply the neighbor based probability classification only after finding large connected components. This approach prevents undesired hole filling, but as a trade-off it rejects small foreground objects. The trade-off may be acceptable in the background replacement application with close-up and medium shots that Li and Sezan are targeting. But the loss of small objects is not desirable in security applications, especially if objects are small because of their distance to the camera. The ability to detect objects with a greater range in sizes requires less zoom to successfully detect objects, thus fewer cameras need be installed to cover the same field of view [3]. The approach of Li and Sezan shows that spatial proximity of pixels assists in classifying them as foreground or background, but it is not sufficient without defining a further relationship between the pixels.

The most obvious relationship between pixels is based on the visual characteristics of the pixels, such as color. Such relationships are complex, e.g. because of texture, and also computationally expensive. This approach depends very much on the progress in still image segmentation.

3SAMMI system overview

Block diagram of the SAMMI system.
Block diagram of the SAMMI system.

The SAMMI (Spatio-Activity Multi-Mode with Iterations) method presented in this paper, makes use of temporal characteristics to define the relationship between neighboring pixels, or blocks of pixels in our particular implementation. The temporal characteristics are cheap to maintain and compare from a resource usage perspective, yet they are effective because they make use of the temporal coherence of moving objects. In this section, we present a general overview of the system, as depicted in figure ?. The key components are discussed in more detail in section ?.

The system context has significant impact on the SAMMI algorithm. The starting point is not raw image data, but a Motion JPEG stream (intra-frame coding only) captured by a camera and transmitted over a network. A departure from conventional approaches is that we don’t decompress the JPEG information to pixel-level information in the spatial domain. Instead, processing is done on Discrete Cosine Transformation (DCT) coefficients in the frequency domain and, as a consequence, at 8x8 pixels block-level. The primary motivation for using the DCT coefficients is the computational cost of the inverse transform to the pixel domain. In addition, it allows for reduction in memory usage and processing for scene modeling. Although storing 192 DCT coefficient values is no different from storing 192 RGB pixels for an 8x8 block, the information carried by the DCT coefficients is concentrated in just a few coefficients. We found that using 8 of the 192 coefficients is sufficient for acceptable detection accuracy. The disadvantage of the block-level approach is that object edges are less accurate than in a pixel-level approach. The impact of the blockiness of edges depends on the application, e.g. security applications are invariant under blockiness, and on the camera resolution.

The original input to our system is in the YCbCr color space. Principal component analysis of the YCbCr component values of typical images shows that Cb and Cr are more highly correlated to each other than the alternative in-phase (I) and quadrature (Q) chrominance components in the YIQ color space [5]. Improving the independence of DCT components improves the results of the DCT classifier described in section ?, allowing for greater invariance under changing lighting conditions such as shadows. The first 6 Y DCT coefficients are used in conjunction with the I and Q DC coefficients.

Using the input coefficients, the scene is modeled on a per-block basis. For each block, several mode models are maintained, each representing a state of the image, representing stationary objects and the underlying background simultaneously. Mode models contain DCT coefficient values and temporal information. This temporal information allows the age and frequency of encountering a mode to be determined. For each mode, the following temporal characteristics are recorded:

  • Creation frame: the frame in which the mode first was created. We refer to the difference between the current frame and the creation frame as the mode age.

  • Hit count: the number of times that the mode has been matched. This characteristic represents the activity of the mode.

  • Last matched frame: the frame that the mode was last matched in.

  • Potential removal frame: the frame in which a mode is scheduled for removal in, if no further matches to this mode occur.

The best mode match for each block in the image is determined by a DCT classifier. The DCT classifier also determines if a new mode is created for a given DCT block. A spatial classifier, described in more detail in section ?, is then applied to improve the selection of matching modes based on the temporal information in adjacent modes.

The temporal information and DCT coefficient values of matched modes are updated using an approximated median filter as described in section ?.

The output of this system is a set of connected components that represent foreground objects. Flood fill connected component analysis is used, using the creation time of the matched mode for each block as the merging criterion. A block is connected to an adjacent block if and only if the creation frame of both matched mode models is either greater than a threshold value T, or smaller or equal to the threshold value T. Therefore, each connected component will be exclusively comprised of blocks that are either entirely of age greater than T or less than or equal to T. Figure ? shows an example where a dropped bag has an intermediate age, compared to the person and the background.

Detecting an abandoned bag: Input image (left), Ages of matched modes in a sequence (right) Detecting an abandoned bag: Input image (left), Ages of matched modes in a sequence (right)

4SAMMI method

4.1DCT classifier

Each block in each input image is compared to the mode models for that block in order to give the probability that the input image block matches the mode model. A high probability of a match means that the input is likely to have been encountered before. If the highest probability mode match for a block is below a threshold T, the input block is likely to have not been encountered previously and a new mode is created. The highest probability mode match can be used directly for creating blob output. However, we employ further processing steps as described in sections ? and ?.

Values for \kappa for an example image. Scene model (left), input image(center), classifier output (dark = matcibh, bright = no match) (right) Values for \kappa for an example image. Scene model (left), input image(center), classifier output (dark = matcibh, bright = no match) (right) Values for \kappa for an example image. Scene model (left), input image(center), classifier output (dark = matcibh, bright = no match) (right)

As an initial mode matching value, a weighted sum of the absolute differences between each of the coefficients of block and mode model is calculated as follows:

where is the difference between the i-th coefficient of the input block and the i-th coefficient of the mode model and is the trained weight. An example of values for an image is shown in Figure ?.

Trained weight is determined as follows. Naive Bayes training is used to determine the probability that a given set of input coefficients match a given set of mode coefficients based upon a set of training data. A logistic regression is then applied to determine coefficients .... This yields the linear regression coefficient of for all .

ROC curve of DCT blocks for match versus no match classification
ROC curve of DCT blocks for match versus no match classification

The threshold for creating a new mode must also be determined. That is, the threshold that determines when a probability is not strong enough to warrant a mode match. We use the criterion {true positive rate + false positive rate = 1} to determine the threshold level for where P(mode match) = 0.5. This can be determined and visualized by using a Receiver Operating Characteristic (ROC) graph (Figure ?). This method is also used for spatial training as detailed in section ?.

4.2Spatial classifier

Video sequences from network cameras often contain noise from camera refocusing, lighting changes and sensor noise. To avoid accumulation of errors in the scene model, a spatial classifier is included in this system. The spatial classifier takes not only the current block into account, but also its neighbors. The underlying assumption is that for a given DCT block at a given point in time in an image sequence, a mode is more likely to be a match if the adjacent DCT blocks match modes that were created at a similar time to when that mode was created.

Indoor scene: Empty scene (top left), system input (top right), system output for 0 spatial iterations (bottom left), system output for 3 spatial iterations (bottom right) Indoor scene: Empty scene (top left), system input (top right), system output for 0 spatial iterations (bottom left), system output for 3 spatial iterations (bottom right) Indoor scene: Empty scene (top left), system input (top right), system output for 0 spatial iterations (bottom left), system output for 3 spatial iterations (bottom right)

Accumulation of noise in the scene model is particularly detrimental to age information. We model the creation frame and hit count of each mode in each block of the sequence. Incorrect mode matching causes the age information to be invalid. These errors can propagate with time due to mode model updating and mode removal, see figure ? (bottom left) for an example. The spatial classifier corrects the DCT classifier result for each block in the sequence by using information from adjacent blocks.

Let adjacent DCT blocks to a given DCT block in an image be defined using 4-connectivity as the blocks directly above, below, left and right of the given DCT block. For each mode model in the scene model for a given DCT block, is the number of mode models of adjacent blocks that are temporally similar to . Temporal similarity is defined according to the difference in creation time between and the adjacent block’s mode model, using threshold t. Value affects the likelihood that is the best match.

The final mode match value for a block is + , where accesses a lookup table for the -th iteration and the number of similar neighbors . The lookup table is constructed during an offline training stage. For each possible value of and , an ROC graph is determined and is taken at the point on the graph where the sum of the true positive rate and false positive rate equals 1. This method is similar to the method for determining the threshold level for the DCT classifier described in section ?. An example of one of the training sequences, at different stages of iteration is shown in figure ?.

Training sequence images, from left to right: Scene model (top left), foreground (top), 0 spatial iterations (top right), 1 spatial iteration (bottom left), 2 spatial iterations (bottom), 3 spatial iterations (bottom right) Training sequence images, from left to right: Scene model (top left), foreground (top), 0 spatial iterations (top right), 1 spatial iteration (bottom left), 2 spatial iterations (bottom), 3 spatial iterations (bottom right) Training sequence images, from left to right: Scene model (top left), foreground (top), 0 spatial iterations (top right), 1 spatial iteration (bottom left), 2 spatial iterations (bottom), 3 spatial iterations (bottom right) Training sequence images, from left to right: Scene model (top left), foreground (top), 0 spatial iterations (top right), 1 spatial iteration (bottom left), 2 spatial iterations (bottom), 3 spatial iterations (bottom right)

4.3Active mode bonus

Mode persistence is used to improve classification. While some object detection applications focus on tracking moving objects, other applications have a greater need for stable and consistent detection of stationary objects. By including a probability measure that is added to the match probability for modes seen within the last few frames, this trade-off can be adjusted by users of the system. Low (or zero) contributions from mode persistence result in better detection of moving objects. Increasing the mode persistence probability results in more stable and consistent stationary object detection, which also reduces the impact of noise eroding a stationary object.

4.4Approximated median filter scene model updating

Exponential Moving Average versus Approximated Median Filter
Exponential Moving Average versus Approximated Median Filter

The commonly used method of using an exponential moving average (EMA) to model scene modes [6] suffers from a number of problems. There is a noise increase in very dark and very light regions, due to sensor noise, lighting changes and foreground objects perturbing the model. Further, the amount of time required to return to the original state after an impulse signal is dependent upon the size of the impulse. In addition, the method is computationally expensive due to the multiplications and precision required.

We use an approximated median filter (AMF) [2] to address the aforementioned issues:

If

If

If and

where is the input coefficient value, is the mode model coefficient value, and is the learning parameter that determines how fast the mode model adapts.

Response to impulse noise for a dark background pixel
Response to impulse noise for a dark background pixel

Biased noise is decreased in bright and dark regions, since the recovery time is dependent upon the duration of the noise signal but not upon the intensity of the noise, see figure ? for an example. Recovery time for impulsive noise is 1 frame, unlike the exponential moving average.

4.5Mode removal

Mode models have to be removed for two reasons. The first reason is practical: a system may run out of memory if too many modes have been created. This is especially important in systems where the system is allocated a fixed maximum amount of memory, e.g. in the context a bigger system where a large number of cameras is supported. In addition, more modes means more processing power is needed. A maximum number of modes may be introduced to make the system performance feasible and predictable. The second reason is regardless of the availability of system resources. Modes must be removed from the system in order to reduce the probability of new objects being matched to unrelated mode models.

Modes that should have been deleted earlier from frequent traffic in the hallway. Scene model (left), input image(center), age image (black = new, green = old), (right) Modes that should have been deleted earlier from frequent traffic in the hallway. Scene model (left), input image(center), age image (black = new, green = old), (right) Modes that should have been deleted earlier from frequent traffic in the hallway. Scene model (left), input image(center), age image (black = new, green = old), (right)

Determining when to remove a mode is a trade-off decision. If modes are removed too soon, objects that are occluded will not match the scene model when they reappear. This is known as the revealed background problem. Modes that are not removed quickly enough have an increased likelihood of being matched to unrelated objects, see for example figure ?. We remove a mode model when it hasn’t been matched to an incoming block for a number of frames. For each mode model, a potential removal frame is computed when the mode model is created. The value is updated each time the mode model is matched to an incoming block. The mode removal frame is computed as follows:

,

where is a constant indicating the minimum survival time of the mode model, and is a constant indicating the minimum percentage of time this part of the scene must remain visible to retain its temporal information.

Each time a new frame is processed, the value of a mode model is checked. If the value matches the current frame number , the mode model is removed regardless of the total number of mode models for the block. Note that always at least one mode model remains if is greater than 0. In addition, each time a new mode model is created, and the maximum number of modes per DCT block has been reached, the mode with the earliest mode removal frame value is removed.

5Evaluation

5.1Experimental setup

An evaluation video data set was recorded using a network camera providing a Motion-JPEG stream (resolution 768x576). Although the SAMMI algorithm is not specific to any particular flavor of JPEG compression, by using a specific camera we have full control over the recording circumstances, and we avoid any influence of video transcoding. The evaluation data set targets difficult, yet realistic scenarios. That is, the data set is typically not representative for the content an object detection system may encounter in practice. The focus of our evaluation is on indoor scenarios, where there is artificial lighting with in some cases additional natural lighting through windows. The data set is biased towards stationary foreground objects, stressing the importance of abandoned object detection and object removal applications. Each video starts with an empty scene to allow for initialization of the background model. See table ? for details on the videos. The screen savers present in some of the videos serve to stretch the modeling in that area of the scene for research purposes, as they have more states than the maximum number of modes.

For each of the 10 videos a bounding box ground truth was created. In addition, for more precise evaluation for in total 10,185 frames in 6 videos, an object mask ground truth was created, which is expected to be accurate within an error margin of 4 pixels. Note that the creation of ground truth is a problem in itself, and that different people or the same person at a different time may segment the same video differently. As creation of the precise object mask ground truth is very expensive manual labor, it was done for a subset of the videos only.

The DCT classifier and spatial classifier are trained on 12 frames acquired from 7 videos. The scenes and locations at which the training videos were recorded are different from the scenes used in the test set.

Description of videos used in the test set. The number of frames used in object mask ground truth is given in square brackets.
Description Difficulty Length (frames) 1. Several people walk through reception area.

5.2Evaluation criteria

Like other object detection algorithms, the SAMMI algorithm has general applicability. Whether the produced output is good in a relative or absolute sense depends on the context in which it is used. The requirements for object detection in an intruder alert system are very different from those in a people counting application. Similarly, a system that alerts a security guard will give a higher penalty to false alarms than a system that does event-based recording. We evaluate the system output at two levels:

  • Pixel-level accuracy: exact matching between ground truth object masks and the detected objects. Note that the output of the SAMMI method has a coarser granularity than pixel, viz. 8x8 blocks. Hence, it is not possible for our method to score the maximum on this level, while pixel-based algorithms could theoretically reach a score of 100%. Also, the problem of inconsistency in ground truths mentioned before may not even allow a perfect segmentation algorithm to score 100%.

  • Tracking suitability: the impact of the segmentation results on a tracker application.

For pixel-level accuracy, we determine a true positive count by an AND operation on the detection result and the ground truth. Similarly, false positives and false negatives are found. An F1-measure (harmonic mean of recall and precision) is then computed.

For tracking suitability, the number of detected blobs is set out against the number of blobs that could be related to a bounding box. Only one blob is associated per ground truth bounding box, resulting in a penalty for fragmentation of objects. Because of the simplicity of the measure, it does not allow us to draw any conclusions about the absolute results. However, the measure does tell to what extent fragmentation has been reduced. Although tracking applications may be able to deal with fragmentation, lower fragmentation leads to lower computational demands and decreases the chance of tracking spurious objects. The tracking suitability measure should always be used in combination with another object detection measure, such as the pixel-level F1-measure, as it does not penalize for missed detections.

Finally, we measure the computational expense and the memory usage of the scene model. As the foremost purpose of the SAMMI method is an increase in speed, it is important to measure the impact on resource usage. Although actual implementations of the evaluated methods have not been optimized for speed, processing time still gives an indication of the difference between methods. Experiments were run on one 2.16 GHz Intel Core Duo processor, and measurements include file I/O. Memory usage is estimated for the persistent scene model only, that is the memory usage in between processing 2 frames.

5.3Results

We compare the SAMMI object detection method for different numbers of iterations to the Mixture of Gaussians method. For SAMMI, the maximum number of mode models was set to 5. Results are shown in table ?. Per video, results for SAMMI using 3 iterations are shown in figure ?

Results for methods according to the evaluation measures.
Method Mixture of Gaussians Number of iterations - 0 1

The Mixture of Gaussian method suffers from the trade-off between detecting moving objects and detecting stationary objects. Although it has a precision higher than SAMMI (0.79 versus 0.69), its recall is significantly lower (0.42 versus 0.95) indicating a large number of false negatives. The difference in performance at pixel level is not significant for the various iterations. However, the larger number of iterations shows a significant increase in tracking suitability, corresponding to a decrease in object fragmentation.

Since our experimental setup does not focus on measuring processing time precisely, it is inappropriate to draw final conclusions about the difference between the methods. However, the significant difference between Mixture of Gaussians and our method supports the expectation that our method is significantly faster. Similarly, the difference in memory usage for scene modeling is significant.

SAMMI performance, using 3 iterations, at pixel level and tracking level for videos for which an object mask ground truth is available.
SAMMI performance, using 3 iterations, at pixel level and tracking level for videos for which an object mask ground truth is available.

6Conclusions

Computationally inexpensive background modeling can be done without a significant penalty in accuracy. The use of DCT information without transforming image information to the pixel domain still allows for good accuracy while making significant savings in resource usage. The use of a fast approximated median method makes the modeling robust to noise in bright and dark regions of a scene, while it is faster than the conventional exponential moving average approach. Fragmentation noise is reduced by several iterations of neighbor adapted classification based on temporal coherency of objects.

Another advantage of the SAMMI system is its configurability. Users can configure the trade-off between detecting new moving objects and existing stationary objects using the active mode bonus. Similarly, users can make trade-offs for removing modes by specifying the minimum percentage of time a part of the scene must remain visible to retain its temporal information.

The spatial processing outlined in this paper allows for a greater variability in the size of objects, particularly small objects, that can be successfully detected. The filtering of local noise in the image sequence that would otherwise cause spurious blobs to be detected is embedded within the scene modeling process.

Through low resource usage while preserving acceptable accuracy, the lightweight object detection method presented in this paper increases the feasibility of deploying video analysis systems in the real world.

References

  1. Adaptive video background replacement.
    B. Li and M. I. Sezan. In Proc Int Conf Multimedia Expo (ICME), pages 385–388, Aug 2001.
  2. Segmentation and tracking of piglets in images.
    N. McFarlane and C. Schofield. Machine Vision and Applications, 8(3):187–193, 1995.
  3. Moving object segmentation using super-resolution background models.
    J. Migdal, T. Izo, and C. Stauffer. In OMNIVIS, Oct 2005.
  4. Automatic vehicle image extraction based on spatio-temporal image analysis.
    T. Nakanishi and K. Ishii. In Proc ICPR, volume I, pages 500–504, Aug 1992.
  5. Digital image processing (2nd ed.).
    W. K. Pratt. John Wiley & Sons, Inc., New York, NY, USA, 1991.
  6. Learning patterns of activity using real-time tracking.
    C. Stauffer and E. Grimson. IEEE Trans Pattern Analysis Machine Intelligence, 22(8):747–757, 2000.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
""
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
   
Add comment
Cancel
Loading ...
21448
This is a comment super asjknd jkasnjk adsnkj
Upvote
Downvote
""
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters
Submit
Cancel

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test
Test description