We study performance characteristics of convolutional neural networks (CNN) for mobile computer vision systems. CNNs have proven to be a powerful and efficient approach to implement such systems. However, the system performance depends largely on the utilization of hardware accelerators, which are able to speed up the execution of the underlying mathematical operations tremendously through massive parallelism. Our contribution is performance characterization of multiple CNN-based models for object recognition and detection with several different hardware platforms and software frameworks, using both local (on-device) and remote (network-side server) computation. The measurements are conducted using real workloads and real processing platforms. On the platform side, we concentrate especially on TensorFlow and TensorRT. Our measurements include embedded processors found on mobile devices and high-performance processors that can be used on the network side of mobile systems. We show that there exists significant latency-throughput trade-offs but the behavior is very complex. We demonstrate and discuss several factors that affect the performance and yield this complex behavior.
Hazardous situations may easily be caused by limited visibility at urban traffic intersections due to buildings, fences, flora and other obstacles. Thus, drivers approaching an intersection have limited reaction time when other obscured road users, such as pedestrians and cyclists, appear unexpectedly. Previous research has been conducted on applications warning drivers of approaching out-of-sight vehicles. However, less focus has been on the detection and awareness applications revealing the presence of pedestrians. We propose a novel system that displays the driver real-time locations and types of hidden road users at traffic intersections. A roadside unit is installed in the infrastructure which sends safety-critical object data to the vehicle, supporting the real-time decision-making of the driver. The roadside unit consists of a monovision camera streaming video to a computing unit which performs object detection and distance measurements on the detected objects. This paper validates the capability of the proposed system of localising a pedestrian, and also examines its sensitivity to installation and detection errors. The results show that the accuracy of the proposed system is suitable for the intended application. However, an error in the vertical angle of the roadside unit camera caused an exponential error in the distance approximation in respect to the measured distance. The detection accuracy was noticed to decrease at long distances and in dark surroundings. Moreover, in order to reduce the effect of the presented errors, the camera should be installed as high as possible without hindering its detection capabilities.
This paper presents a dynamically reconfigurable multi-radio RF architecture concept, which can be used for RF platform and control optimization. The platform realization is based on the RF hardware and its configuration mechanisms. The related control software is realized through functional separation of configuration management and timing control. Both key hardware and software elements are discussed and optimization opportunities evaluated using high-level analysis on key building blocks.
Multi-radio platforms are an interesting design concept. Executing multiple radios on a shared platform presents opportunities not only for component re-use but also for better data throughputs, as non-active radios may dynamically yield resources for active ones. This enhances the conventional SDR approach in the RF domain and provides means to optimize resources in platform level when taking the link and network traffic issues into account. Such flexibility can provide opportunities for future cognitive radios when operating in heterogeneous networks. The downside is increased RF interference, and thus, receiver desensitization. We review the design and performance trade-offs of multi-radio platforms focusing on LTE and WLAN and present motivation for simple co-operation mechanisms to their future revisions.
We study methods for observing physical defects on the surface of power cables. Quality control is essential for power cable manufacturing and surface defects are an important quality factor. Traditionally power cable manufacturing has relied on manual inspection as automated methods have not been sufficient to be used in industrial production.We have designed and implemented a novel defect detection system that applies machine learning methods to detect power cable surface defects. Our system uses laser scanning to map the surface of a cable during production. For the machine learning, we have evaluated different CNN (Convolutional Neural Network) architectures and studied their performance and accuracy. According to our results, CNNs are suitable for the detection of surface defects as they can be trained with large amounts of cable surface data.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.