Gesture Recognition in Vehicles

[0001] This application is a continuation of U.S. patent application Ser. No. 13/762,332 filed Feb. 7, 2013 to El Dokor et al. titled Three-Dimensional Gesture Recognition in Vehicles, currently pending, which is in turn is a continuation of U.S. patent application Ser. No. 12/784,022 filed May 20, 2010 to El Dokor et al. titled Systems and Related Methods for Three-Dimensional Gesture Recognition in Vehicles, now U.S. Pat. No. 8,396,252, the entire contents of each of these applications thereof being incorporated herein by reference.

[0002] This invention relates generally to implementations of three dimensional (3D) gesture recognition systems and related methods, and more particularly, relate to communication and data access systems for use in vehicles.

[0003] Particular implementations of 3D gesture recognition systems and related methods can enable data access and communication by a vehicle driver while maximizing the time the driver is watching the road. Because the use of gesture recognition permits the driver to interact with a computing system without requiring touching or use of an interface device (such as a keyboard or mouse), the driver's hands are essentially free at almost all times for performing driving tasks. Other implementations of 3D gesture recognition systems and related methods may be used in a wide variety of other systems and use conditions. Promoting eyes on the road is a key aspect of any vehicular human-machine interface.

[0004] Conventional approaches to 3D gesture recognition exist. Conventional approaches to 3D gesture recognition suffer from a number of difficulties such as robustness, perspective, lens aberrations, pose, weak responses to lighting conditions, sensor saturation, weak performance under very violent vibration conditions, differences in skin tone from user to user, variable lighting conditions, and many other challenges. Conventional solutions typically approach solving these problems by developing predictive algorithms that utilize definite and absolute thresholds for various critical parameters in their algorithms. The problem with such systems is that they cannot easily adapt if they encounter situations that are beyond a particular algorithm's threshold, or outside the scope of usability of a given set of algorithms and implementations. Other approaches utilize time-of-flight cameras to mitigate the need for scene reconstruction, or even creating a depth map. The problem with TOF sensors is that they have very unpredictable responses to both, very strong ambient lighting conditions, as well as very high vibration-motion blur is a major issue with TOF sensors. Motion blur can wreak havoc on the image acquisition process, undermining the entire gesture recognition system.

[0005] Therefore, it would be desirable to provide an apparatus that overcomes the drawbacks of the prior art.

[0006] In accordance with one or more embodiments of the present invention, implementations of 3D gesture recognition systems are described that combine machine learning methodologies with vision algorithms for TOF camera-based systems to create an adaptive capability that permits the system to modify its algorithms to “learn” each user of the system. Because the system moves from a set of default, pre-trained values for its algorithms to a set of learned values without using absolute thresholds, the system does not rely on external algorithm revisions to improve its accuracy. The system can improve its own algorithms as it interacts with the user. Implementations of 3D gesture recognition systems described in this document are sufficiently robust for use in accuracy-critical industries like the automotive industry and others where adaptive capability is critical if gesture recognition systems are to be widely adopted.

[0007] Still other objects and advantages of the invention will in part be obvious and will in part be apparent from the specification and drawings.

[0008] The invention accordingly comprises the several steps and the relation of one or more of such steps with respect to each of the others, and the apparatus embodying features of construction, combinations of elements and arrangement of parts that are adapted to affect such steps, all as exemplified in the following detailed disclosure, and the scope of the invention will be indicated in the claims.

[0017] Various embodiments of the invention will now be described making reference to the figures in which like reference numbers denote like steps or structure. Referring first to FIG. 1, an implementation of a 3D gesture recognition system in accordance with an embodiment of the present invention is illustrated. As is shown in FIG. 1, a first embodiment of a 3D gesture recognition system may include a time of flight (TOF) camera 110 coupled with a computing platform 120. The computing platform may include or may be coupled with one or more databases 130. Output from the computing platform may be displayed on a two dimensional (2D) screen or display 140 that may or may not include touch interaction capability in particular implementations. The TOF camera is oriented relative to a vehicle control 150 present in the vehicle so that gestures made by one or both hands of a driver operating the vehicle, passenger in the vehicle, or other desired individual in the vehicle may be acquired as a series of images. In the implementation illustrated in FIG. 1, the vehicle control is preferably a steering wheel. In other implementations, the vehicle may comprise a train, airplane or the like, and thus the vehicle control will be presented as appropriate for the particular vehicle in use.

[0018] Examples of TOF cameras that may be utilized in particular implementations of 3D gesture recognition systems that may be constructed in accordance with one or more embodiments of the invention may include systems manufactured by Canesta, Centre Suisse d'Electronique et de Microtechnique SA (CSEM), or 3DV Systems. While these cameras may all differ in their capabilities and properties, they all can be used in conjunction with a depth engine incorporated in the computing system.

[0019] Examples of 2D displays that may be utilized in particular implementations of 3D gesture recognition systems constructed in accordance with one or more embodiments of the invention may include an ultra-lightweight organic LED screen, as well as more conventional screens currently being utilized in vehicles. In particular implementations, a 3D display could also be utilized.

[0020] The computing platform may include any of a wide variety of components, including one or more processors, memory, one or more cards, controllers, electrically erasable programmable read-only memory (EEPROM), power supplies, fans, and any other electronic component. In particular implementations, an embedded video card system may be utilized.

[0021] During operation of the system, images collected by TOF camera 110 are sent to computing platform 120 for processing. During the processing step computing platform 120 may utilize the one or more databases 130 to determine on or more gestures the series of images may represent. Computing platform 120 is adapted to correlate any recognized gestures with one or more predefined commands and then to implement the commands by changing an aspect of the user interface image displayed by the 2D screen (causing the user interface image to scroll, a new image to appear, or indicating the pressing of an object within the image), and many other user interface activities, or other commands or activities associated with the operation or other aspects of the vehicle in question.

[0022] A static gesture is a particular orientation of the hand or arm or other gesture indicator that may be employed by a user, such as a pointer or the like. Static gestures include orientations of the hand or arm that are recognized when the hand or arm forms a pattern that does not include a movement (such as many American Sign Language signs). Dynamic gestures include orientations of the hand or arm that are recognized when the hand, fingers, palm, wrist, or arm move in a particular orientation, perform a defined action, or make a defined movement. Implementations of 3D gesture recognition systems presented in accordance with one or more embodiments of the present invention are preferably configured to recognize both static and dynamic gestures, but may be configured to recognize only one or the other, or alternatively may be configurable as desired to recognize one, the other, or both types of gestures. Without using a combination of both static and dynamic gestures, however, implementations of 3D gesture recognition systems may lack the ability to operate effectively, and while they may in fact be compromised in one or more ways, they may also be applicable on one or more particular situations.

[0023] Generally, and referring to FIG. 2, implementations of methods of 3D gesture recognition presented in accordance with an embodiment of the invention preferably comprise five basic steps: Preprocessing 210, Pre-gesture recognition (depth map, face detection, etc. . . . ) 220, Gesture recognition 230, Interface implementation 240, and Algorithm parameter adjustment/learning 250. Each of these steps may contain a number of other processing steps, and each will be considered in turn. Each of, or any combination of these steps may be implemented in or as part of an implementation of a depth engine included within the computing system.

[0024] During preprocessing step 210, and referring to FIG. 3 which depicts the preprocessing step in greater detail, implementations of 3D gesture recognition methods in accordance with an embodiment of the present invention include may include various steps at the preprocessing stage. As illustrated, the computing system may receive raw depth data included in raw, or unprocessed, frame data from the TOF camera at step 310. TOF cameras generally provide a pixel by pixel depth map of any objects in the field of view in contrast with video or other cameras that provide a pixel by pixel light reflectance map of the objects in the field of view. The raw frames from the TOF camera may include influences from a wide variety of factors, including, by non-limiting example, lens aberration, user skin tone variation, lighting conditions, distance of the camera, foreshortening, camera alignment, signal aliasing, or any other factor influencing a TOF camera image.

[0025] As is further illustrated in FIG. 3, implementations of 3D gesture recognition systems may remove, compensate, adjust, or correct for any of the foregoing factors by utilizing implementations of methods of cleaning the depth data of each frame at step 320. A wide variety of conventional, calculated or predetermined, filtering techniques may be utilized in particular implementations of the method of one or more embodiments of the invention. When the data has been cleaned, the images may be normalized, enhanced, or otherwise improved to allow the system to observe the gestures made with greater accuracy and robustness.

[0026] At step 330 of FIG. 3, the cleaned depth data from the TOF camera may then be subjected to 3D (frame-by-frame) and 4D (frame-to-frame with time being the fourth dimension) clustering to determine any objects of interest contained in the data. At this stage, the objects of interest may include the body of the user, other major appendages, or any other object (such as a portion of the vehicle) within the field of view that may or may not be used in future processing. A wide variety of statistical, algorithmic, parametric, or other methods may be utilized to perform the clustering of the cleaned depth data in particular implementations of 3D gesture recognition methods. One or more particular methods of performing the clustering may be implemented using a wide variety of computer languages and data processing techniques within the computing system, including, by non-limiting example, MATLAB®, C++, C#, Visual Basic, Compute Unified Device Architecture (CUDA), or any other programming language or combination of computing or programming techniques.

[0027] Once the cluster(s) corresponding with one or more objects of interest have been identified, in particular implementations of 3D gesture recognition system algorithms in accordance with embodiments of the invention, portions of the depth data not corresponding with the objects of interest may be disregarded and not processed in subsequent steps. In other implementations, some or all of this data may continue to be included. In either implementation, various implementations of clustering techniques are utilized to identify from the cluster data containing the objects of interest which cluster corresponds with the body of the user (the body cluster) at step 340.

[0028] Once the body cluster has been identified, implementations of the method of one or more embodiments of the invention may then use various implementations of clustering techniques to find from the clusters corresponding with the objects of interest the cluster(s) corresponding with the arm(s) of the user at step 350. In particular implementations, the clustering technique implementations may utilize searching algorithms that look for the arm clusters in specific regions relative to the body cluster. After the arm clusters have been identified, the specific clustering technique implementations utilized in 3D gesture recognition method implementations may then locate the head and shoulders at step 360. In particular implementations, the head and shoulders may be located as segments from the clusters corresponding to the objects of interest; in other embodiments, the location of the clusters may be calculated and then identified from the objects of interest; and in still other embodiments, the head and shoulders may be located by searching in specific regions defined relative to the body cluster. Since during use conditions, the user's head and/or shoulders may be angled relative to the TOF camera, a wide variety of correction techniques may be used to resolve the angle and eliminate it from the images, or otherwise compensate therefore to improve system performance and/or accuracy. An alternate approach would be to train a deep belief network on various orientations, such that the output (label neurons) identify gestures with a predefined set of degrees of freedom.

[0029] Once the head and/or shoulders have been identified, implementations of 3D gesture recognition systems in accordance with embodiment of the invention at step 370 may calculate the arm length of the user using previously learned and/or stored information that relate the arm length of an individual to his or her height, shoulder proportion, head size, or any other relevant measurement of a human male or female. Once the calculation has been performed, implementations of 3D gesture recognition systems in accordance with one or more embodiments of the present invention may perform a search using implementations of clustering techniques in an area of the “cleaned” frame data that corresponds with the calculated arm length, may perform a search within a range of distances within the cleaned frame data relative to the head or other located portion of the body, or use any other method of searching for cluster data corresponding with the palm of the hand at step 380. Because arm length proportions vary from person to person, and particularly for children and teenagers, a set of pre-defined or learned parameters may be used that may be adaptively changed based on actual data received regarding the user. An alternative would be to trace the depth values back to real-world coordinates, and then use data biometrics for looking up or calculating such parameters. The adaption may take place as the parameters are adjusted in real-time based on failures or other errors during image processing, or the adaption may take place in a feed-back fashion after each or a certain number or type of gestures have been recognized and/or processed.

[0030] Referring to FIG. 3, once the palm of the hand has been identified, implementations of 3D gesture recognition systems in accordance with one or more embodiments of the invention may calculate the tip of the hand at step 390. In particular implementations the tip of the hand may be defined as the location of the tip of the middle finger, or other predetermined and/or desirable portion thereof. With the position of the palm and the tip of the middle finger identified, the system is now ready for pre-gesture recognition steps.

[0031] Referring next to FIG. 4, during pre-gesture recognition, implementations of 3D gesture recognition methods presented in accordance with one or more embodiments of the present invention may define a set of predetermined search spaces within each frame of cleaned depth data in which to look for the tip of the middle finger and palm or look for movement of the tip of the middle finger and palm at step 410. Within each of these search spaces, one or more partition spaces may be defined at step 420 that divide the search spaces into smaller areas. Then, at step 430, when the tip of the finger/palm/hand or any other portion of the hand enters one of the partition spaces, foveation may be used in particular implementations to model the pixels containing the cleaned depth data to maintain highest resolution at the tip of the finger or other desired focus point on the hand and assign lower resolution to some or all of the other clusters in each frame image (like the body cluster). By utilizing foveation, the computing power of the computing system may be utilized most effectively as the computing system may need to process only portions of each frame image at its highest resolutions. This may also improve accuracy, robustness, and learning in particular implementations.

[0032] Once the tip of the finger/other portion of the hand has been identified, then implementation of 3D gesture recognition systems presented in accordance with one or more embodiments of the present invention may determine whether the hand portion is tilted vertically relative to the camera, toward the camera, or away from the camera at step 440. Once the tilt has been identified, the system may utilize various algorithms at step 450 to resolve the tilt, such as, by non-limiting example, principal component analysis, or any other method of determining the orientation of an object in a set of depth data. If the tilt is not removed from the data, the number of gesture recognition algorithms may be needlessly increased to ensure that each tilted gesture is recognized the same as any non-tilted gesture. The level of accuracy, robustness, and repeatability of the system may also be adversely affected in particular implementations.

[0033] Referring once again to FIG. 4, once the tilt of the hand has been resolved, at step 460 previously learned information may be used to locate the hand with a certain level of certainty. The level of certainty may be determined using statistical or experimental methods and the learned information may be stored in the databases accessible to the computing system.

[0034] Once the hand has been located with a certain level of certainty, implementations of 3D recognition systems presented in accordance with one or more embodiments of the present invention may now be ready to perform various gesture recognition steps. Referring next to FIG. 5, implementations of 3D recognition systems presented in accordance with one or more embodiments of the invention may execute a context-aware algorithm at step 510. Implementations of context-aware algorithms may include algorithms that retrieve from a table of known gestures those that are active (or allowed for execution) in the particular context of the gesture just observed. For example, when the image on the 2D screen is centered over an object on the map, the only two gestures that may be active may be a gesture that allows for selection of the object or a gesture that moves the image away from being centered over the object. All other gestures may be ignored, even though implementations of 3D recognition systems may attempt to recognize some or all gestures the user is performing while driving or otherwise operating the vehicle in question.

[0035] Once a gesture has been observed and the specific active gesture algorithms retrieved by the computing system, implementations of 3D gesture recognition systems may determine whether the gesture observed is a static or dynamic gesture at step 520. This determination may be made by performing a first comparison of the gesture with any of the active static and/or dynamic gestures. In other implementations, the determination may be made by analyzing each frame or group of frames for certain elements that identify a static or dynamic gesture (such as speed or direction of hand movement, or position of a particular portion of the hand). If at step 520 it is determined that the gesture was a static gesture, processing continues at step 530 and the observed gesture may be matched with a set of known masks included in the one or more databases using an algorithm for each known static gesture (images may be represented by numbers in a table, or full images of all 1000 possible hand positions stored in a lookup table, for example) using a cross correlation technique or other image correlation technique to make the comparison. If the gesture was identified at step 520 as a dynamic gesture, processing passes instead to step 540, and features from clips of the observed gesture may be extracted and matched with motion streams stored in the one or more databases. A deep belief network is used for gesture recognition. Input data representing the hand are used with the associated DBN. A neural network or other video or image correlation technique may be employed to make the correlation. Particular implementations of gesture recognition systems may utilize 10-15 frame segments of the observed gestures during the feature extraction and correlation steps.

[0036] Once the observed gesture has been recognized, implementations of 3D gesture recognition systems presented in accordance with one or more embodiments of the present invention are now ready to implement one or more commands corresponding with the recognized gesture using the computing system and/or the 2D display. Referring to FIG. 6, the implementation may first implement step 610 for retrieving a command corresponding with the matched gesture that is to be implemented on the 2D display using the current context of the gesture and/or the nature of the gesture itself, and at step 620 executing the command using the computing system and/or the 2D display. An indication that the command has been or is being executed may be visible on the 2D display or may be indicated by audible or other feedback to the user (push the button, call on the telephone, scroll the map, raise the bubble, etc). In addition to presenting a response on the 2D display, other responses may be possible in response to the issues command, including one or more vehicle actions.

[0037] During or after each or any number of gesture recognition process completions or after or during any step within the gesture recognition process, implementations of 3D gesture recognition system presented in accordance with one or more embodiments of the present invention may employ a wide variety of machine learning algorithms and/or process implementations that permit the system and/or method to be adjusted to more accurately recognize gestures made by a particular user or users. Referring finally to FIG. 7, a high level diagram of the basic steps included in this wide variety of machine learning algorithms and/or process implementations is illustrated. As is shown, embodiments of the present invention may include at step 710 reviewing a previously executed gesture recognition process or completed processes or the just executed step or steps and determining any of a wide variety of sub-optimal conditions such as, by non-limiting example, errors, retries, time-outs, warning conditions, data corruption problems, or any other less desirable characteristic of the process and/or step. At step 720, new values for algorithm/system parameters may be determined using information obtained during the review. These new parameters may be stored at step 730 in the one or more databases or the computing system itself. Because machine learning algorithms and or/process implementations are utilized in implementations of 3D recognition systems and methods, the resulting systems and methods may be capable of evolving to accurately capture specific user gestures and may avoid the need for external algorithm modification, thus allowing the system and methods to be adaptively used and implemented.

[0038] One or more of the 3D gesture recognition systems presented in accordance with one or more of the embodiments of the present invention may have one or more of the following advantages over the current state-of-the-art. The use of clustering algorithms, along with foveation, permit more robust and faster recognition of gestures. The ability of implementations of the system to adapt its own recognition algorithms permits it to become more robust over time without requiring external intervention and eliminating errors. The ability of the system to use the body of the user as a reference eliminates the requirement that the user maintain an exact position or be still for a predetermined period of time before gesture recognition can take place. Many other advantages exist and will be evident to those of ordinary skill in the art using the principles disclosed in accordance with the embodiments of this invention.

[0039] The materials used for the computing systems, TOF cameras, 2D displays, and databases may be made of conventional materials used to make goods similar to these in the art, such as, by non-limiting example, plastics, metals, semiconductor materials, rubbers, composites, and ceramics. Those of ordinary skill in the art will readily be able to select appropriate materials and manufacture these products from the disclosures provided herein.

[0040] The implementations listed here, and many others, will become readily apparent from this disclosure. From this, those of ordinary skill in the art will readily understand the versatility with which this disclosure may be applied.

[0041] Still other objects and advantages of the invention will in part be obvious and will in part be apparent from the specification and drawings.

[0042] The invention accordingly comprises the several steps and the relation of one or more of such steps with respect to each of the others, and the apparatus embodying features of construction, combinations of elements and arrangement of parts that are adapted to affect such steps, all as exemplified in the following detailed disclosure, and the scope of the invention will be indicated in the claims.