Abstract: Various examples are directed to a System on a Chip (SOC) and methods of operating the same. The SOC may access firmware code from a boot Read-Only Memory (ROM) of the SOC. The firmware code may comprise a plurality of functional blocks. The SOC may determine that firmware patching is active for the SOC and access patch data from a non-volatile memory of the SOC. The SOC may determine that a first functional block of the firmware code has been patched and access first patch code from the patch data. The first patch code may be associated with the first functional block. The SOC may execute the first patch code.

Abstract: This disclosure describes techniques to perform convolutional neural networks (CNNs) on embedded devices. The techniques include operations comprising: accessing DNN information including definition of layers and weights of the DNN; obtaining cache or memory information for one or more cache or memory levels of the resource constrained embedded device; and configuring the DNN to be loaded onto the one or more cache or memory levels of the resource constrained embedded device based on the cache or memory information and the DNN information.

Abstract: This disclosure describes techniques to aggregate sensor data. The techniques perform operations comprising: receiving, from a first sensor, a first profile representing a first set of movement attributes detected by the first sensor in an area at a given point in time; receiving, from a second sensor, a second profile representing a second set of movement attributes detected by the second sensor in the area at the given point in time; computing a similarity measure between the first and second sets of movement attributes of the first and second profiles; determining that the similarity measure exceeds a threshold value; and in response to determining that the similarity measure exceeds the threshold value, associating the first and second profiles with a same first object that is in the area at the given point in time.

Abstract: A method for fine-tuning a convolutional neural network (CNN) and a sensor system based on a CNN are disclosed. The sensor system may be deployed at a deployment location. The CNN may be fine-tuned for the deployment location using sensor data, e.g., images, captured by a sensor device of the sensor system at the deployment location. The sensor data may include objects that are not present in an initial data set used for training the CNN. The sensor data and the initial data set may be input to the CNN to train the CNN and obtain fine-tuned parameters of the CNN. The CNN can thus be fine-tuned to the deployment location of the sensor system, with an increased chance of recognizing objects when using the sensor system and the CNN to recognize objects in captured sensor data.

Abstract: Various examples are directed to a System on a Chip (SOC) and methods of operating the same. The SOC may access firmware code from a boot Read-Only Memory (ROM) of the SOC. The firmware code may comprise a plurality of functional blocks. The SOC may determine that firmware patching is active for the SOC and access patch data from a non-volatile memory of the SOC. The SOC may determine that a first functional block of the firmware code has been patched and access first patch code from the patch data. The first patch code may be associated with the first functional block. The SOC may execute the first patch code.

Abstract: This disclosure describes techniques to count objects across multiple zones of arbitrary shapes. The techniques include operations comprising: detecting an object in a first zone of a plurality of zones of an area; determining that the object has moved from the first zone to a second zone of the plurality of zones; determining that movement of the object from the first zone to the second zone fails to satisfy a zone transition criterion for updating a count value and time criterion; and in response to determining that the movement of the object from the first zone to the second zone fails to satisfy the zone transition criterion, preventing the count value from being updated.

Abstract: This disclosure describes techniques to detect an object. The techniques include operations comprising: receiving an image captured by overhead camera; identifying a region of interest (ROI) of a plurality of regions within the image; selecting an object classifier from a plurality of object classifiers based on a position of the identified ROI relative to the overhead camera; and applying the selected object classifier to the identified ROI; and detecting presence of the object within the ROI in response to applying the selected object classifier to the identified ROI.

Abstract: This disclosure describes techniques to count objects in an area. The techniques include operations comprising: receiving, from a first sensor, data that identifies a location of a first object in an area at a given point in time; receiving, from a second sensor, data that identifies a location of a second object in the area at the given point in time; and aggregating the data from the first sensor and the data from the second sensor to compute a count value indicating how many objects are present in the area at the given point in time.

Abstract: A method for fine-tuning a convolutional neural network (CNN) and a sensor system based on a CNN are disclosed. The sensor system may be deployed at a deployment location. The CNN may be fine-tuned for the deployment location using sensor data, e.g., images, captured by a sensor device of the sensor system at the deployment location. The sensor data may include objects that are not present in an initial data set used for training the CNN. The sensor data and the initial data set may be input to the CNN to train the CNN and obtain fine-tuned parameters of the CNN. The CNN can thus be fine-tuned to the deployment location of the sensor system, with an increased chance of recognizing objects when using the sensor system and the CNN to recognize objects in captured sensor data.

Abstract: This disclosure describes techniques to perform convolutional neural networks (CNNs) on embedded devices. The techniques include operations comprising: accessing DNN information including definition of layers and weights of the DNN; obtaining cache or memory information for one or more cache or memory levels of the resource constrained embedded device; and configuring the DNN to be loaded onto the one or more cache or memory levels of the resource constrained embedded device based on the cache or memory information and the DNN information.

Abstract: One factor in limiting the speed of conventional implementations of mixture models is that the algorithm involves many decisions where different operations are fetched and performed depending on the outcome of the decisions. These decisions cause flushing of the pipeline, and thus prevent the realization of a highly parallel pipeline in a processor. Without parallelism, the throughput of the pipeline in the processor, i.e., the ability to process many samples of the digital input at a time, is limited. To alleviate this issue, implementation of the mixture model is reformulated, among other things, by embedding decisions into the process flow as multiplicative factors. The resulting implementation alleviates the need to use if-else statements for the decisions and reduces the number of times the pipeline has to be flushed. The implementation enables a pipeline with a higher degree of parallelism and thereby increases throughput and speed of the implementation.

Abstract: This disclosure describes techniques to count objects across multiple zones of arbitrary shapes. The techniques include operations comprising: detecting an object in a first zone of a plurality of zones of an area; determining that the object has moved from the first zone to a second zone of the plurality of zones; determining that movement of the object from the first zone to the second zone fails to satisfy a zone transition criterion for updating a count value and time criterion; and in response to determining that the movement of the object from the first zone to the second zone fails to satisfy the zone transition criterion, preventing the count value from being updated.

Abstract: This disclosure describes techniques to aggregate sensor data. The techniques perform operations comprising: receiving, from a first sensor, a first profile representing a first set of movement attributes detected by the first sensor in an area at a given point in time; receiving, from a second sensor, a second profile representing a second set of movement attributes detected by the second sensor in the area at the given point in time; computing a similarity measure between the first and second sets of movement attributes of the first and second profiles; determining that the similarity measure exceeds a threshold value; and in response to determining that the similarity measure exceeds the threshold value, associating the first and second profiles with a same first object that is in the area at the given point in time.

Abstract: This disclosure describes techniques to count objects in an area. The techniques include operations comprising: receiving, from a first sensor, data that identifies a location of a first object in an area at a given point in time; receiving, from a second sensor, data that identifies a location of a second object in the area at the given point in time; and aggregating the data from the first sensor and the data from the second sensor to compute a count value indicating how many objects are present in the area at the given point in time.

Abstract: This disclosure describes techniques to detect an object. The techniques include operations comprising: receiving an image captured by overhead camera; identifying a region of interest (ROI) of a plurality of regions within the image; selecting an object classifier from a plurality of object classifiers based on a position of the identified ROI relative to the overhead camera; and applying the selected. object classifier to the identified ROI; and detecting presence of the object within the ROI in response to applying the selected object classifier to the identified ROI.

Abstract: One factor in limiting the speed of conventional implementations of mixture models is that the algorithm involves many decisions where different operations are fetched and performed depending on the outcome of the decisions. These decisions cause flushing of the pipeline, and thus prevent the realization of a highly parallel pipeline in a processor. Without parallelism, the throughput of the pipeline in the processor, i.e., the ability to process many samples of the digital input at a time, is limited. To alleviate this issue, implementation of the mixture model is reformulated, among other things, by embedding decisions into the process flow as multiplicative factors. The resulting implementation alleviates the need to use if-else statements for the decisions and reduces the number of times the pipeline has to be flushed. The implementation enables a pipeline with a higher degree of parallelism and thereby increases throughput and speed of the implementation.

Abstract: Many conventional video processing algorithms attempting to detect human presence in a video stream often generate false positives on non-human movements such as plants moving in the wind, rotating fan, etc. To reduce false positives, a technique exploiting temporal correlation of non-human movements can accurately detect human occupancy while reject non-human movements. Specifically, the technique involves performing temporal analysis on a time-series signal generated based on an accumulation of foreground maps and an accumulation of motion map and analyzing the running mean and the running variance of the time-series signal. By determining whether the time-series signal is correlated in time, the technique is able to distinguish human movements and non-human movements. Besides having superior accuracy, the technique lends itself to an efficient algorithm which can be implemented on low cost, low power digital signal processor or other suitable hardware.

Abstract: System and techniques for a visual vehicle parking occupancy sensor are described herein. A color image, including a parking space, is received from a camera. A cascaded search for vehicle features in a hue-saturation-value (HSV) converted version of the color image is performed to produce search results. A search for macro vehicle features in the color image is also performed to produce an indication of found macro vehicle features when the search results are of a first type. An occupancy indicator is provided based on the search results when the search results are of a second type and based on the indication otherwise.

Abstract: System and techniques for a visual vehicle parking occupancy sensor are described herein. A color image, including a parking space, is received from a camera. A cascaded search for vehicle features in a hue-saturation-value (HSV) converted version of the color image is performed to produce search results. A search for macro vehicle features in the color image is also performed to produce an indication of found macro vehicle features when the search results are of a first type. An occupancy indicator is provided based on the search results when the search results are of a second type and based on the indication otherwise.

Abstract: A system and method for removing noise from images are disclosed herein. An exemplary system includes an edge-detection-based adaptive filter that identifies edge pixels and non-edge pixels in an image and selects a filtering technique for at least one non-edge pixel based on a comparison of the at least one non-edge pixel to a neighboring pixel region, wherein such comparison indicates whether the at least one non-edge pixel is a result of low-light noise.