Search

Numerous vehicle detection methods have been proposed to obtain trustworthy traffic data for the development of intelligent traffic systems. Most of these methods perform sufficiently well under common scenarios, such as sunny or cloudy days, however, the detection accuracy drastically decreases under various bad weather conditions, such as rainy days or days with glare, which normally happens during sunset. This study proposes a vehicle detection system with a visibility complementation module that improves detection accuracy under various bad weather conditions. Furthermore, the proposed system can be implemented without retraining the deep learning models for object detection under different weather conditions. The complementation of the visibility was obtained through the use of a dark channel prior and a convolutional encoder&ndash, decoder deep learning network with dual residual blocks to resolve different effects from different bad weather conditions. We validated our system on multiple surveillance videos by detecting vehicles with the You Only Look Once (YOLOv3) deep learning model and demonstrated that the computational time of our system could reach 30 fps on average, moreover, the accuracy increased not only by nearly 5% under low-contrast scene conditions but also 50% under rainy scene conditions. The results of our demonstrations indicate that our approach is able to detect vehicles under various bad weather conditions without the need to retrain a new model.

Dynamic voltage and frequency scaling (DVFS) is a well-known method for saving energy consumption. Several DVFS studies have applied learning-based methods to implement the DVFS prediction model instead of complicated mathematical models. This paper proposes a lightweight learning-directed DVFS method that involves using counter propagation networks to sense and classify the task behavior and predict the best voltage/frequency setting for the system. An intelligent adjustment mechanism for performance is also provided to users under various performance requirements. The comparative experimental results of the proposed algorithms and other competitive techniques are evaluated on the NVIDIA JETSON Tegra K1 multicore platform and Intel PXA270 embedded platforms. The results demonstrate that the learning-directed DVFS method can accurately predict the suitable central processing unit (CPU) frequency, given the runtime statistical information of a running program, and achieve an energy savings rate up to 42%. Through this method, users can easily achieve effective energy consumption and performance by specifying the factors of performance loss.

In this paper, Haar-like patterns and a classification machine were employed for fingerprint feature extraction. Although the Gabor filter has high accuracy for feature extraction under a wide angle representation range, the full range of the filter is not required. Moreover, fingerprints have clear direction field representation. Due to these two key factors, this paper uses modified Haar-like patterns to create near-circular patterns for acceptable feature detection. A support vector machine classifier is used to detect features. In a performance comparison that used 15 general Haar-like patterns as a benchmark, the proposed algorithm was able to reduce the average computation required by approximately 50% without sacrificing accuracy. Thus, the proposed method is suitable for real-world applications.

Most object detection models cannot achieve satisfactory performance under nighttime and other insufficient illumination conditions, which may be due to the collection of data sets and typical labeling conventions. Public data sets collected for object detection are usually photographed with sufficient ambient lighting. However, their labeling conventions typically focus on clear objects and ignore blurry and occluded objects. Consequently, the detection performance levels of traditional vehicle detection techniques are limited in nighttime environments without sufficient illumination. When objects occupy a small number of pixels and the existence of crucial features is infrequent, traditional convolutional neural networks (CNNs) may suffer from serious information loss due to the fixed number of convolutional operations. This study presents solutions for data collection and the labeling convention of nighttime data to handle various types of situations, including in-vehicle detection. Moreover, the study proposes a specifically optimized system based on the Faster region-based CNN model. The system has a processing speed of 16 frames per second for 500 &times, 375-pixel images, and it achieved a mean average precision (mAP) of 0.8497 in our validation segment involving urban nighttime and extremely inadequate lighting conditions. The experimental results demonstrated that our proposed methods can achieve high detection performance in various nighttime environments, such as urban nighttime conditions with insufficient illumination, and extremely dark conditions with nearly no lighting. The proposed system outperforms original methods that have an mAP value of approximately 0.2.