With a focus on practical applications in the real world, a number of challenges impede the progress of pedestrian detection. Scale variance, cluttered backgrounds and ambiguous pedestrian features are the main culprits of detection failures. According to existing studies, consistent feature fusion, semantic context mining and inherent pedestrian attributes seem to be feasible solutions. In this paper, to tackle the prevalent problems of pedestrian detection, we propose an anchor-free pedestrian detector, named context and attribute perception (CAPNet). In particular, we first generate features with consistent well-defined semantics and local details by introducing a feature extraction module with a multi-stage and parallel-stream structure. Then, a global feature mining and aggregation (GFMA) network is proposed to implicitly reconfigure, reassign and aggregate features so as to suppress irrelevant features in the background. At last, in order to bring more heuristic rules to the network, we improve the detection head with an attribute-guided multiple receptive field (AMRF) module, leveraging the pedestrian shape as an attribute to guide learning. Experimental results demonstrate that introducing the context and attribute perception greatly facilitates detection. As a result, CAPNet achieves new state-of-the-art performance on Caltech and CityPersons datasets.
This article focuses on a two-hop relay system, where a potential malicious single-antenna relay may forward altered symbols to a base station (BS) equipped with massive antennas, while the BS also receives signals from direct channel between itself and the source. The main challenge is that the malicious relay may not only alter symbols, but also transmit misleading pilot symbols to impair the channel estimation at the BS. With this regard, we investigate attack detection without reliable channel estimation. During each channel coherence interval, the BS receives pilot signal, namely, reference vector (RV), and immediately examines reliability of the RV by utilizing its massive antennas. If it is reliable, the received information signals, namely, information vectors, will be projected into the RV. The projection gives compressed signal with one dimension. After $n$ intervals, the BS could obtain an $n$ -length sequence of compressed signals across $n$ different channel coherence intervals and an $n$ -length sequence of direct signals from the direct channel. We prove that arbitrary attacks are detectable by comparing empirical distribution of these two sequences with its a priori stochastic distribution. Furthermore, simulation confirms that our proposed schemes outperform other existing attack detection methods.
With a focus on integrated sensing, communication, and computation (ISCC) systems, multiple sensor devices collect information of different objects and upload it to data processing servers for fusion. Appearance gaps in composite images caused by distinct capture conditions can degrade the visual quality and affect the accuracy of other image processing and analysis results. The authors propose a fused‐image harmonisation method that aims to eliminate appearance gaps among different objects. First, the authors modify a lightweight image harmonisation backbone and combined it with a pretrained segmentation model, in which the extracted semantic features were fed to both the encoder and decoder. Then the authors implement a semantic‐related background‐to‐foreground style transfer by leveraging spatial separation adaptive instance normalisation (SAIN). To better preserve the input semantic information, the authors design a simple and effective semantic‐aware adaptive denormalisation (SADE) module. Experimental results demonstrate that the authors’ proposed method achieves competitive performance on the iHarmony4 dataset and benefits from the harmonisation of fused images with incompatible appearance gaps.