Automated measurement of mouse social behaviors using depth sensing, video tracking, and machine learning

Most current mouse tracking software is designed for use with 2D videos recorded from a top- or side-view camera (24–28). Two-dimensional video analysis has several limitations, such as difficulty resolving occlusion between animals, difficulty detecting vertical movement, and poor animal tracking performance against backgrounds of similar color. To overcome these problems, we developed an integrated hardware setup with synchronized image acquisition and software to record behavior using synchronized video cameras and a depth sensor. Depth sensors detect depth values of an object in the z-plane by measuring the time-of-flight of an infrared light signal between the camera and the object for each point of the image (32), in a manner analogous to sonar.

We compared two commercially available depth sensors, the Kinect Sensor from Microsoft Corporation and the Senz3D depth and gesture sensor from Creative Technology Ltd. (Fig. S1). We developed customized software to acquire raw depth images from both sensors. Although the Kinect sensor has been recently used for behavioral tracking in rats (31), pigeons (33), pigs (34), and human (35), we found that its spatial resolution was not sufficient for resolving pairs of mice, which are considerably smaller; in contrast, the Senz3D sensor’s higher 3D resolution made it better suited for this application (Fig. S1). This higher resolution was partly because the Senz3D sensor was designed for a closer working range (15–100 cm) than the Kinect (80–400 cm). In addition, the Senz3D sensor’s smaller form factor allowed us to build a compact customized behavior chamber with 3D video acquisition capability and space for insertion of a standard mouse cage (Fig. 1A).

We installed a side-view conventional video camera in front of the cage as well as a top-view video camera and the Senz3D sensor on top of the cage (Fig. 1 A and B). Videos taken from the side-view and top-view cameras provided additional and complementary data, such as luminosity, for the postacquisition image analysis and behavior analysis and allowed users to manually inspect and score behaviors from different angles. Data were acquired synchronously by all three devices to produce simultaneous depth information and top- and side-view grayscale videos (Methods). Representative video frames from each of the three devices during three social behaviors (aggression, mounting, and close investigation) are shown in Fig. 1C.

Mice are nocturnal animals, and exposure to white light disrupts their circadian cycle. Therefore, we recorded animal behaviors under red light illumination, which is considered “dark” for mice, because mice cannot perceive light within the red-to-infrared spectrum. Both video cameras and the depth sensor in our system work under red light and do not rely on white-light illumination. To separate animal identities, our system is currently limited to tracking and classifying two mice of different coat colors.

The major steps of the postacquisition image analysis and behavior analysis (Fig. 1D and Fig. S2) are described in the following sections.

To integrate the monochrome video recordings from the top-view camera with the data from the depth sensor, we registered them into a common coordinate framework using the stereo calibration procedure from MATLAB’s Computer Vision System Toolbox, in which a planar checkerboard pattern is used to fit a parameterized model of each camera (Fig. 2 A and B and Fig. S3). The top-view camera and depth sensor were placed as close as possible to each other, to minimize parallax (Fig. 2A). We then projected the top-view video frames into the coordinates of the depth sensor (Fig. S3) to obtain simultaneous depth and intensity values for each pixel.

We performed background subtraction and image segmentation using reconstructed data from the top-view camera and depth sensor to determine the location and identity of the two animals (Fig. S4 and Methods). To obtain a low-dimensional representation of animal posture (“pose”), we fit an ellipse to each animal detected in the segmented video frames (Fig. 2C). The body orientation of each animal was determined from its position and movement direction, as well as from features detected by a previously developed machine learning algorithm (24, 27) (Methods). Thus, the pose of each animal is described by a set of five parameters from the fit ellipse: centroid position (x, y), length of the major axis (l), length of the minor axis (s), and body orientation (θ). To evaluate the performance of the automated pose estimation, we constructed a ground truth dataset of manually annotated ellipses and calculated the differences between automatic and manual estimation for the parameters including centroid position, body length, head orientation, and head position (Fig. 2D). We also evaluated the overall performance by computing the weighted differences between the machine annotated and manually annotated ellipses using a previously developed metric (27) (Fig. 2D and Methods). We found that the algorithm was able to track the position and the body orientation of the animals in a robust manner (Fig. 2D and Movie S1), compared with the performance of two independent human observers (Fig. S5).

Using the five fit ellipse parameters and additional data from the depth sensor, we developed a set of 16 second-order features describing the state of each animal in each video frame (Fig. 3 A and B) and 11 “window” features computed over multiple frames, giving 27 total features (Methods). Principal component analysis of features extracted from a sample set of movies indicated that the 16 second-order features are largely independent of each other, such that the set of recorded behaviors spanned nearly the entire feature space (Fig. 3 C and D).

We next explored the use of supervised machine learning approaches for automated annotation of social behavior. In supervised learning, classifiers are trained using datasets that have been manually annotated with the desired classifier output, to find a function that best reproduces these manual annotations. The performance of the classifier is evaluated using a testing set of ground-truth videos not used for training. The training set and the test set have no overlap and were obtained from separate videos. We used our 27 extracted features to test several supervised learning algorithms, including support vector machine (SVM), adaptive boosting (adaBoost), and random decision forest (TreeBagger). The random decision forest gave us the best performance in prediction accuracy and training speed and was thus selected for further investigation. We trained three social behavior classifiers (attack, mounting, and close investigation; see Methods for the criteria used by human annotators) using a set of six videos that contained ∼150,000 frames that were manually annotated on a frame-by-frame basis. We chose to generate 200 random decision trees, which was beyond where the error rate plateaued (Fig. 4G); because individual decision trees were built independently, the process of training the decision forest is parallelizable and can be greatly sped up on a multicore computer. The output of our three behavior detectors for three representative videos is shown in Fig. 4 A–D (male–male interactions) and Fig. 4 E and F (male–female interactions). As seen in the expanded raster plots (Fig. 4 B, D, and F), there is a qualitatively close correspondence between ground truth and prediction bouts for attack, close investigation, and mounting. The contribution of individual features to classifier performance is shown in Fig. 4H.

To measure the accuracy of these behavior classifiers in replicating human annotations, we manually labeled a set of 14 videos (not including the videos used to train the classifier) that contained ∼350,000 frames from a variety of experimental conditions and measured classifier error on a frame-by-frame basis. We plotted classifier performance using the detection error tradeoff (DET) curve representing the framewise false negative rate vs. the false-positive rate (Fig. 4I) and the precision-recall curve representing the framewise true positive rate vs. the positive predictive rate (Fig. 4J), using the human annotations as ground truth. These measurements illustrated the tradeoff between the true positive rate vs. the positive predictive value at different classification thresholds from 0 to 1. Here we chose a classification threshold that optimized the framewise precision and recall; the framewise precision, recall, fallout, and accuracy rates at the classification threshold are shown in Fig. 4K. All of the classifiers showed an overall prediction accuracy of 99% for attack, 99% for mounting, and 92% for close investigation. Finally, we measured the precision and recall rates at the level of individual behavioral episodes (“bouts”), periods in which all frames were labeled for a given behavior. We observed a high level of boutwise precision and recall across a range of minimum bout durations (Fig. 4K and Movies S2–S4).

To explore the utility of our behavior classifiers, we used them to track several biologically relevant behaviors under several experimental conditions. We first used the classifier to annotate resident male behavior during interactions with either a male or a female intruder (Fig. 5; I_m vs. I_f, respectively). We examined the percentage of time resident males spent engaging in attack, mounting, and close investigation of conspecifics (Fig. 5 A–C); note that this parameter is not directly comparable across behaviors, because the average bout length for each behavior may be different. We therefore also measured the total numbers of bouts during recording (Fig. 5 D–F), the latency to the first bout of behavior for each resident male (Fig. 5 G–I), and the distribution of bout lengths for each behavior (Fig. 5 J–R). We observed that for our standard strain C57BL/6N, male residents (R_C57N) exhibited more close investigation bouts with longer duration toward male (Fig. 5N; I_m) than that toward female (Fig. 5K; I_f) intruders (P < 0.001), although the total numbers of bouts were comparable between the two conditions (Fig. 5E). The classifier predictions showed no significant differences from the ground truth in the measured percentage of time spent engaging in each behavior, nor in the bout length distribution of each behavior (Fig. 5 K, N, and Q, yellow vs. gray bars) (∼350,000 frames total), suggesting that the same classifiers work robustly in both male–male and male–female interactions.

To examine how genetic backgrounds influence social behaviors, we compared two strains of resident male mice, C57BL/6N and NZB/B1NJ (Fig. 5). NZB/B1NJ mice were previously shown to be more aggressive than C57BL/6N (36). Consistently, we found that NZB/B1NJ resident males spent more time attacking BALB/c intruder males, and significantly less time engaging in close investigation, than did C57BL/6N resident males (Fig. 5 A and B; R_NZB) (P < 0.05). This likely reflects a more rapid transition from close investigation to attack, because the average latency to attack was much shorter for NZB/B1NJ than for C57BL/6N males (Fig. 5G). Interestingly, NZB/B1NJ animals exhibited both a higher number of attack bouts (Fig. 5D) (P < 0.05) and longer average attack durations compared with C57BL/6N animals (Fig. 5 M and P) (P < 0.05). These data illustrate the ability of the method to reveal differences between the manner in which NZB/B1NJ and C57BL/6N males socially interacted with intruder animals of a different strain. In all measurements, the classifier prediction showed no significant differences from the ground truth (Fig. 5), suggesting that the same classifiers work robustly with distinct strains of animals that exhibit very different social behaviors.

To explore the utility of our behavioral classifiers in detecting social deficits in mouse models of autism, we examined the behavior in Black and Tan Brachyury (BTBR) T+tf/J (BTBR) mice, an inbred mouse strain that was previously shown to display autism-like behavioral phenotypes, such as reduced social interactions, compared with C57BL/6N animals (1, 37–39). Here we measured parameters of social interactions between BTBR mice (or C57BL/6N control mice) and a “target” animal of the BALB/c strain, in an unfamiliar, neutral cage. By using our behavioral analysis system to track the locations, poses, and behaviors of the interacting animals, we observed significantly less social investigation performed by BTBR animals in comparison with C57BL/6N controls (Fig. 6 A–C), consistent with previous reports (38, 39). In particular, the BTBR animals displayed shorter bouts of (Fig. 6B), and reduced total time engaged in (Fig. 6C), social investigation.

To determine whether this reduction of social investigation reflects less investigation of the BALB/c mouse by the BTBR mouse (in comparison with the C57BL/6N controls), or vice versa, we measured the social investigation behavior performed by the BALB/c mouse. BALB/c animals did not exhibit reduced social interactions with the BTBR mice in comparison with C57BL/6N controls (Fig. S6 A and B). This suggests that the reduction of social investigation observed in BTBR animals is indeed due to less investigation of the BALB/c mouse by the BTBR mouse.

Finally, we asked whether pose estimation and supervised behavioral classifications offered additional information beyond tracking animal location alone. We first measured “body–body” distance—the distance between centroid locations of two interacting animals (illustrated in the schematic in Fig. 6D)—a measurement that only used the output from tracking animal location alone but not from pose estimation or behavioral classifiers. We observed a trend to decreased time spent at short body–body distances (<6 cm) in BTBR animals (Fig. 6 D and E), but this effect was not statistically significant. When we measured “head–body” distance—the distance between the front end of the subject and the centroid of the other animal (illustrated in the schematic in Fig. 6F)—a measurement that used output from both tracking and pose estimation, but not from supervised behavioral classifications, we observed a statistically significant reduction in time spent at short (<4 cm) head–body distances in BTBR animals paired with BALB/c mice (Fig. 6 F and G), compared with that in C57BL/6N animals paired with BALB/c. This difference did not reflect reduced investigation of BTBR animals by BALB/c mice, because the latter did not show a significant difference in time spent at short head–body distances toward BTBR vs. C57BL/6N mice (Fig. S6 C and D). Rather, the difference reflects reduced close investigation of BALB/c mice by BTBR mice in comparison with C57BL/6N controls. These data together suggest that our behavioral tracking system was able to detect social behavioral deficits in BTBR mice, a mouse model of autism, and that compared with animal location tracking alone, pose estimation and supervised behavioral classification provide additional useful information in detecting behavioral and phenotypic differences.

The customized behavioral chamber was designed in SolidWorks 3D computer-aided design software and manufactured in a machine shop at California Institute of Technology. To obtain depth images of animals, a depth sensor (Creative Senz3D depth and gesture camera) was installed on top of the cage. The Creative Senz3D depth and gesture camera is a time-of-flight camera that resolves distance based on the speed of light, by emitting infrared light pulses and measuring the time-of-flight of the light signal between the camera and the subject for each point of image. Because mice are nocturnal animals but are insensitive to red light, animal behaviors are recorded under red light illumination (627 nm). Although the depth sensor itself contains an RGB color camera, it does not perform well under red light illumination; therefore, we installed a monochrome video camera (Point Gray Grasshopper3 2.3 MP Mono USB 3.0 IMX174 CMOS camera) on top of the cage, next to the depth sensor. The top-view camera and depth sensor were placed as close as possible to each other, to minimize parallax, and were aligned perpendicular to the bottom of the cage. To obtain side-view information, an additional monochrome video camera of the same model was installed in front of the cage.

Frames from the two monochrome video cameras were recorded by StreamPix 6.0 from NorPix Inc. Similar software did not exist for the Senz3D camera, and therefore we developed customized C# software to stream and record raw depth data from the depth sensor. All three devices were synchronized by customized MATLAB scripts. Frames of each stream and their corresponding timestamps were recorded at 30 frames per second and saved into an open source image sequence format (SEQ) that can be accessed later in MATLAB using a MATLAB Computer Vision toolbox (vision.ucsd.edu/∼pdollar/toolbox/doc/).

To integrate the monochrome information from the top-view camera with the depth information from the depth sensor, we registered the top-view camera and the depth sensor into a common coordinate frame using the Stereo Calibration and Scene Reconstruction tools from MATLAB’s Computer Vision System Toolbox (included in version R2014b or later), in which a calibration pattern (a planar checkerboard) was used to fit a parameterized model of each camera (Fig. S3 A–D). We then projected the top-view video frames into the coordinates of the depth sensor (Fig. S3 E–H) to obtain simultaneous depth and monochrome intensity values for each pixel.

We performed background subtraction and image segmentation using data from top-view and depth sensor to determine the location and identity of the two animals (e.g., resident and intruder in the resident-intruder assay or two individuals in the reciprocal social interaction assay). Specifically, we first determined the rough locations of animals in each frame using image segmentation based on the depth information. The 3D depth of the unoccupied regions of the mouse cage was stitched together from multiple frames to form the depth background of the entire cage. This background was then subtracted from the movie to remove objects in the mouse’s home environment, such as the water dispenser. We then performed a second round of finer-scale location tracking, in which the background-subtracted depth images were segmented to determine the potential boundary of the animals, and the identities of the animals were determined by their fur colors (black vs. white) using data from the monochrome camera. Each segmented animal in each frame was fit with an ellipse, parameterized by the centroid, orientation, and major- and minor-axis length. Body orientation was determined using an automated algorithm that incorporated data from the animal’s movement velocity and rotation velocity and from feature outputs produced by a previously developed algorithm (24, 27).

To evaluate the performance of the automated estimation of animal poses, we constructed a ground truth data set of 634 manually annotated ellipses and calculated the frame-wise differences in fit and measured centroid position, length of the major axis (estimated body length), heading orientation, and head position (Fig. 2D):

where ${\theta }_n$ denotes the value of the pose parameter (centroid position, estimated body length, or estimated head position) with n ∈ {1 for measured value, 2 for fit value}, and l is the measured length of the major axis.

We also calculated framewise differences in fit and measured heading orientation:

where $\lfloor x\rfloor =\min \left(x,360-x\right)$ and ${\theta }_n$ denotes the value of the heading orientation with $n\in \left\{1\hspace{0.17em}\text{for}\hspace{0.17em}\text{measured}\hspace{0.17em}\text{value},\hspace{0.17em}2\hspace{0.17em}\text{for}\hspace{0.17em}\text{fit}\hspace{0.17em}\text{value}\right\}$.

We evaluated the overall performance of individual frames by computing the framewise weighted differences between the machine-annotated and manually annotated ellipses using a previously developed metric (27),

where D is the number of pose parameters, ${\sigma }_i^2$ denotes the variance of the differences between human annotations of the ith pose parameter, and ${\theta }_n^i$ denotes the value of the ith pose parameter with $n\in \left\{1\hspace{0.17em}\text{for}\hspace{0.17em}\text{measured}\hspace{0.17em}\text{value},\hspace{0.17em}2\hspace{0.17em}\text{for}\hspace{0.17em}\text{fit}\hspace{0.17em}\text{value}\right\}$.

We also compared the differences between two independent human observers (Fig. S5) and determined the 98% percentile, which was used as the threshold for evaluating the performance of the prediction.

Using the five fit ellipse parameters and additional data from the depth sensor, we developed a set of 16 second-order features describing the state of each animal in each video frame, and 11 “window” features computed over multiple frames, giving 27 total features. These features are described in Supporting Information and Fig. S7.

In supervised learning, classifiers are trained using datasets that have been manually annotated with the desired classifier output, to construct a function that best reproduces these manual annotations. We used the 27 extracted features to test several supervised learning algorithms, including SVM, adaptive boosting (adaBoost), and random decision forests (TreeBagger). The random decision forest gave us the best performance in prediction accuracy and training speed and was thus selected for further investigation. We trained three social behavior classifiers (attack, mounting, and close investigation) using a set of six videos of male–male and male–female interactions, in which a total of ∼150,000 frames were manually annotated on a frame-by-frame basis.

We then trained an ensemble of 200 random classification trees using the TreeBagger algorithm in MATLAB; output of the classifier was taken as the mode of the bagged trees. We chose to use 200 trees for classification because this number was well beyond where error rate plateaued; because trees in the ensemble can be trained in parallel, increasing the size of the population was not computationally expensive.

To measure the accuracy of the three decision forest classifiers in replicating human annotations, we manually labeled a different set of 14 videos from a variety of experimental conditions that contained ∼350,000 frames total, and used them as our test set. We evaluated the performance using the DET curve representing false negative rate $FNR=FN/FN+TP$ and the false positive rate $FPR=FP/FP+TN$, as well as the precision-recall curve representing the true positive rate $TPR=TP/TP+FN$ vs. the positive predictive value $PPV=TP/TP+FP$. We also measured the accuracy by computing the fraction of true positive and true negative in all classes $ACC=TP+TN/TP+TN+FN+FP$, where TP is true positive, TN is true negative, FP is false positive, and FN is false negative.

Boutwise precision was defined as

where ${\displaystyle {\sum }_{i\in TP}{T}_i}$ is the total time of the true-positive bouts and ${\displaystyle {\sum }_{i\in FP}{T}_i}$ is the total time of the false-positive bouts. Here a true-positive bout was the classified bout in which >30% of its frames were present in the ground truth; a false-positive bout was the classified bout in which ≤30% of its frames were present in the ground truth.

Boutwise recall was defined as

where ${\displaystyle {\sum }_{i\in TP}{T}_i}$ is the total time of the true-positive bouts and ${\displaystyle {\sum }_{i\in FN}{T}_i}$ is the total time of the false-negative bouts. Here a true-positive bout was the ground truth bout in which >30% of its frames were present in the classification; a false-negative bout was the ground truth bout in which ≤30% of its frames were present in the classification.

Experimental subjects were 10-wk-old wild-type C57BL/6N (Charles River Laboratory), NZB/B1NJ (Jackson Laboratory), and BTBR T+tf/J (Jackson Laboratory). In the resident-intruder assay to examine attack, social investigation, and mounting behaviors (Figs. 4 and 5), intruder mice were BALB/c males and females, purchased at 10 wk old (Charles River Laboratory). In the reciprocal social interaction assay to examine social investigation (Fig. 6), interaction partners were BALB/c males, purchased at 10 wk old (Charles River Laboratory). The intruder male was gonadally intact and the intruder females were randomly selected. Animals were housed and maintained on a reversed 12-h light-dark cycle for at least 1 wk before behavioral testing. Care and experimental manipulations of animals were in accordance with the NIH Guide for the Care and Use of Laboratory Animals and approved by the Caltech Institutional Animal Care and Use Committee.

The resident-intruder assay was used to examine aggression, mounting, and close investigation of a resident mouse in its home cage (Figs. 4 and 5). Resident males in their home cages were transferred to a behavioral testing room containing a customized behavioral chamber equipped with video acquisition capabilities (described in Video Acquisition and Camera Registration). An unfamiliar male or female (“intruder”) mouse was then introduced into the home cage of the tested resident. The resident and intruder were allowed to interact with each other freely for 15∼30 min before the intruder was removed. If excessive tissue damage was observed due to fighting, the interaction was terminated prematurely.

A reciprocal social interaction assay was used to examine social investigation between two interacting animals in an unfamiliar, neutral cage (Fig. 6). The procedure followed was similar to the resident-intruder assay, except that both interacting individuals were introduced to an unfamiliar, neutral cage and were allowed to interact with each other for 15 min before they were removed.

Two synchronized videos were scored manually on a frame-by-frame basis using a Computer Vision MATLAB toolbox (vision.ucsd.edu/∼pdollar/toolbox/doc/). The human observer performing annotation was blind to experimental conditions. In manual scoring, each frame was annotated as corresponding to aggression, mounting, or close-investigation behavior. Aggression was defined as one animal engaging in biting and tussling toward another animal; mounting was defined as one animal grabbing on to the back of another animal and moving its arms or its lower body; close investigation was defined as the head of one animal closely investigating any body parts of another animal within 1/4 ∼1/3 body length.

Statistical analysis was performed using MATLAB (MathWorks). The data were analyzed using two-sample t test, two-sample Kolmogorov–Smirnov test, and Mann–Whitney U test.

The depth-sensing camera used in this paper, Senz3D, was recently discontinued by Intel Corp. and Creative, Inc. However, an alternative device, DepthSense 325, is being sold by SoftKinetics, Inc. Senz3D and DepthSense 325 are identical products, except for their product packaging and branding. They offer identical functionalities and are supported by an identical software development kit (SDK).