SmartDetector: Intelligent Sensor Detection

• SmartDetector is a family of intelligent detection systems that fuse raw sensor data with task-specific intelligence to overcome domain bottlenecks.
• They integrate multimodal approaches including thermal feature fusion, on-sensor inference, and context monitoring to enhance reliability and efficiency.
• These systems employ architectural innovations and hardware-level processing to improve accuracy and interpretability in complex, real-world environments.

SmartDetector is a recurrent research designation for detector systems that augment raw detection with an additional layer of intelligence: small-object feature fusion in thermal imagery, explicit context or degradation awareness, embedded or on-sensor inference, multimodal hazard sensing, or interpretable similarity analysis. In the literature represented here, the name is applied to heterogeneous systems rather than a single canonical architecture, spanning SSD-derived visual detectors (Zhang et al., 2021), UV imaging flame sensors (Volpe et al., 2017), radar-only human detection in harsh environments (Liu et al., 19 Jan 2026), self-aware visual monitoring under image degradation (Becker et al., 20 Feb 2026), edge-adaptive mobile detection control (Callegaro et al., 2022), intelligent pixel ASICs (Badea et al., 2024), and smart-contract function similarity detection (Liu et al., 11 Sep 2025).

1. Term, scope, and recurrent pattern

The term covers multiple technical lineages whose common feature is not sensor modality, but the addition of task-specific intelligence beyond a plain detector head. In some works this intelligence is architectural, as in dilation–deconvolution feature fusion for small-object detection (Zhang et al., 2021). In others it is system-level, such as environmental condition monitoring in ADAS (Tran et al., 2021), degradation-space self-awareness (Becker et al., 20 Feb 2026), or dynamic edge offloading control (Callegaro et al., 2022). It can also be embodied directly in hardware, as in a 28 nm pixel ROIC with on-chip neural inference (Badea et al., 2024), or in physical sensing substrates such as GEM-based UV flame imagers and multifunctional avalanche gaseous detectors (Volpe et al., 2017, Abbrescia et al., 2020).

Domain	Representative formulation	Primary target
Thermal/small-object vision	DDSSD on SSD300	Small objects in RGB and thermal images
ADAS and self-aware vision	Backbone17-Det with environment classifier; degradation manifolds	Road-scene detection under context shift or degraded imaging
Mobile edge detection	SmartDet with Katch-Up and DRL control	Object detection under bandwidth and latency constraints
4D radar perception	Model-driven radar-only pipeline	Human detection in dust-filled or underground spaces
Hazard sensing	Triple GEM UV imager; SMART IDS	Flames, sparks, smoke, gases, Radon
On-sensor HEP readout	Intelligent pixel ROIC in 28 nm CMOS	High-$p_{\mathrm{T}}$ cluster filtering at 40 MHz
Software analysis	Statement-tree similarity detector	Smart-contract clone detection

This suggests that SmartDetector is best interpreted as a family resemblance term: detectors become “smart” when they explicitly model the operational bottleneck of their domain, whether that bottleneck is missing semantics at shallow layers, visibility degradation, wireless latency, detector bandwidth, or clone interpretability.

2. Vision architectures for small, cluttered, and difficult targets

In thermal and small-object vision, SmartDetector is closely associated with SSD-derived feature engineering. DDSSD is built on SSD300 with a reduced VGG-16 backbone and retains SSD’s multi-scale prediction scheme, but it replaces the raw Conv4_3 detection map with a fused representation constructed from a Dilation Module on Conv4_3 and a Deconvolution Module on fc7 (Zhang et al., 2021). The fusion is written as

$$X^+ = C\{D(X_{\text{Conv4\_3}}), T(X_{\text{fc7}})\},$$

where $D(\cdot)$ enlarges the receptive field at the shallow layer and $T(\cdot)$ upsamples deeper semantic features to the $38\times 38$ resolution. The motivation is explicit: in SSD, small objects are mainly detected from Conv4_3, but that layer lacks semantic information, while deeper layers have richer semantics but lower spatial resolution (Zhang et al., 2021).

The resulting detector preserves SSD’s multi-scale heads and default boxes while improving small-object behavior. DDSSD reports $79.7\%$ mAP on PASCAL VOC2007 test and $28.3\%$ mmAP on MS COCO test-dev at 41 FPS with $300\times 300$ input on a single Nvidia 1080 GPU, and on the FLIR thermal dataset it reports $22.8$ AP overall, $30.1$ AP$$X^+ = C\{D(X_{\text{Conv4\_3}}), T(X_{\text{fc7}})\},$$0, and $$X^+ = C\{D(X_{\text{Conv4\_3}}), T(X_{\text{fc7}})\},$$1 AR$$X^+ = C\{D(X_{\text{Conv4\_3}}), T(X_{\text{fc7}})\},$$2 in the detailed ablation-oriented description (Zhang et al., 2021). The FLIR setting is particularly significant because over $$X^+ = C\{D(X_{\text{Conv4\_3}}), T(X_{\text{fc7}})\},$$3 of person, car, and bicycle instances have area below $$X^+ = C\{D(X_{\text{Conv4\_3}}), T(X_{\text{fc7}})\},$$4 pixels, and the imagery comes from a $$X^+ = C\{D(X_{\text{Conv4\_3}}), T(X_{\text{fc7}})\},$$5 FLIR Tau2 thermal camera with $$X^+ = C\{D(X_{\text{Conv4\_3}}), T(X_{\text{fc7}})\},$$6 daytime and $$X^+ = C\{D(X_{\text{Conv4\_3}}), T(X_{\text{fc7}})\},$$7 night content (Zhang et al., 2021).

A related, but distinct, “smart detector” design appears in SCRDet for small, cluttered, and rotated objects, especially in aerial imagery (Yang et al., 2018). SCRDet combines SF-Net for finer anchor sampling and multi-layer fusion, MDA-Net for supervised pixel attention plus channel attention, and an IoU-modulated Smooth L1 formulation for rotated bounding-box regression. On DOTA, the full system with image pyramids reports $$X^+ = C\{D(X_{\text{Conv4\_3}}), T(X_{\text{fc7}})\},$$8 mAP on the oriented-bounding-box task, while ablation shows progression from $$X^+ = C\{D(X_{\text{Conv4\_3}}), T(X_{\text{fc7}})\},$$9 for the R$D(\cdot)$0CNN baseline to $D(\cdot)$1 before image pyramids (Yang et al., 2018). Here, “smartness” lies in making sampling, attention, and regression geometry explicitly match the failure modes of small and rotated objects.

A more application-specific visual SmartDetector is the smoke-detector recognizer intended for future drone integration in industrial facilities (Kratochvila et al., 25 Mar 2026). That work compares YOLOv11, SSD, and RT-DETRv2 under real and semi-synthetic training strategies, and evaluates on a normal test set and a difficult test set containing motion blur, small resolution, and incomplete objects. The best detector is YOLOv11n with an average [email protected] of $D(\cdot)$2, while SSD with VGG16 reaches $D(\cdot)$3 and exhibits the smallest Test-N/Test-D gap (Kratochvila et al., 25 Mar 2026). On embedded deployment with $D(\cdot)$4 images and a ROS-2 pipeline, YOLOv11n runs at about $D(\cdot)$5 FPS on Raspberry Pi 5, whereas SSD-VGG16 and RT-DETRv2-L are substantially slower (Kratochvila et al., 25 Mar 2026). This establishes a practical vision-oriented SmartDetector pattern: moderate model scale, mixed real and semi-synthetic data, and explicit attention to embedded runtime.

3. Context-aware, self-aware, and edge-adaptive detection

A second lineage defines SmartDetector by its ability to reason about its own operating context rather than only emitting detections. In ADAS, one formulation couples a compact YOLOv2-based one-stage detector, Backbone17-Det, with an environmental condition classifier and a cloud warning mechanism (Tran et al., 2021). Backbone17 introduces residual connections and SE channel-attention modules into a 17-convolution-layer backbone, while the environmental classifier uses grayscale Average Gray Value with empirically chosen thresholds $D(\cdot)$6, $D(\cdot)$7, and $D(\cdot)$8 to distinguish daytime, twilight, nighttime with street light, and nighttime without street light (Tran et al., 2021). On the BDD100K person subset, the detector reports $D(\cdot)$9 AP30 and $T(\cdot)$0 AP50 at $T(\cdot)$1, close to a Faster R-CNN baseline at AP50 $T(\cdot)$2, while the system-level novelty is that significant context change triggers a warning to the cloud rather than being silently ignored (Tran et al., 2021).

A more formalized notion of self-awareness appears in degradation-manifold monitoring (Becker et al., 20 Feb 2026). Here, a lightweight embedding head is attached to multiple backbone layers, and the representation space is trained via contrastive learning so that images with the same degradation composition are close while different degradation configurations are separated. A pristine prototype $T(\cdot)$3 defines the nominal operating point, and the self-awareness signal is the cosine deviation

$T(\cdot)$4

Because this score is image-level and independent of the detector’s box confidences, it addresses silent failure under blur, noise, compression, weather, or resolution change (Becker et al., 20 Feb 2026). On COCO mixed corruptions, the YOLOv10-m backbone reports AUROC $T(\cdot)$5 at severity 1 and $T(\cdot)$6 at severity 5, and cross-dataset zero-shot results remain high on KITTI, DETRAC, UAVDT, VISDRONE, and FLIR (VIS) (Becker et al., 20 Feb 2026).

SmartDet extends context-awareness into the networking and systems layer for mobile object detection (Callegaro et al., 2022). Its architecture runs object tracking locally on a Jetson Nano, offloads selected frames for object detection to an edge server, and applies a Katch-Up mechanism that re-tracks buffered frames once a delayed edge detection result arrives. A DRL controller then chooses the offloading period, the EfficientDet variant, and whether to activate Katch-Up, optimizing

$T(\cdot)$7

with $T(\cdot)$8 (Callegaro et al., 2022). On the real-world Jetson Nano–GTX 980 Ti testbed, SmartDet reports a $T(\cdot)$9 increase in mAR relative to a baseline with full Katch-Up usage and maximum channel usage while using $38\times 38$0 less channel and $38\times 38$1 less power resources associated with Katch-Up; relative to a fixed minimal-resource strategy, it improves mAR by $38\times 38$2 while using Katch-Up on one third of the frames (Callegaro et al., 2022). In this usage, SmartDetector denotes adaptive control over the detector’s computational and communication envelope.

4. Radar and hazard-sensing realizations

In harsh industrial and subterranean settings, SmartDetector is instantiated as a radar-only, model-driven perception pipeline rather than a neural visual detector. The 4D mmWave radar framework for pervasive human detection processes radar points $38\times 38$3 through domain-aware multi-threshold filtering, ego-motion compensated temporal accumulation of two frames, KD-tree Euclidean clustering with Doppler-aware refinement, and rule-based 3D classification (Liu et al., 19 Jan 2026). The system runs on an NVIDIA Jetson Orin NX and is evaluated in a dust-filled enclosed trailer and in underground mining tunnels. On a 160-frame sequence with two pedestrians always present, it reports frame-wise recall $38\times 38$4 ($38\times 38$5), person-count recall $38\times 38$6 ($38\times 38$7), and false-alarm rate about $38\times 38$8 of frames, whereas a camera-based YOLOv8 baseline drops to $38\times 38$9 frame-wise recall and $79.7\%$0 person-count recall (Liu et al., 19 Jan 2026). Here the “smart” element is interpretability and domain-aware tuning under conditions where camera and LiDAR collapse.

A different physical realization is the GEM-based UV imaging flame detector (Volpe et al., 2017). That system uses a sealed chamber with a UV-transparent window, a triple resistive GEM stack, a CsI photocathode coated with a thin ethylferrocene layer, a pad readout plane, a lens, and a narrow-band UV filter. The detector operates in Ne + $79.7\%$1 CH$79.7\%$2 or Ne + $79.7\%$3 CF$79.7\%$4 at 1 atm, employs a reversed drift field of about 200 V, and reaches gas gain $79.7\%$5 (Volpe et al., 2017). It is reported as almost $79.7\%$6 times more sensitive than the best commercial flame sensor in the abstract, and the detailed description states $79.7\%$7 to $79.7\%$8 sensitivity improvement, $79.7\%$9 better time resolution, operation in direct sunlight, and imaging of a candle flame at 15 m with $28.3\%$0 pad readout (Volpe et al., 2017). Because it images rather than only triggers, it can determine where the flame or spark appears.

The SMART integrated detector system generalizes this hazard-sensing idea into a multifunctional node for environmental monitoring (Abbrescia et al., 2020). It targets flames, sparks, smoke, dangerous gases, and Radon using MWPC and RETGEM structures derived from CERN detector technologies. The paper reports sensor characteristics between $28.3\%$1 and $28.3\%$2 times higher than the best commercial devices, flame sensitivity between $28.3\%$3 and $28.3\%$4 higher than EN-54-10 Class-1 detectors, gas sensitivity improvements of about $28.3\%$5 to $28.3\%$6 in avalanche mode relative to simple ionization mode, and stable Radon detection in $28.3\%$7 humid air with efficiencies for alpha particles approaching $28.3\%$8 (Abbrescia et al., 2020). The architecture is explicitly networked: each sensor unit is intended to transmit to a common computing centre via cellular or satellite phone protocols (Abbrescia et al., 2020).

5. Embedded intelligence at the sensor front end

SmartDetector can also mean that the detector itself performs local inference before readout. The intelligent pixel detector for future collider tracking embeds clustering, feature extraction, and neural-network classification into a radiation-hard readout integrated circuit in TSMC 28 nm bulk CMOS (Badea et al., 2024). The chip contains two $28.3\%$9 physical-pixel arrays on a $300\times 300$0 die, with $300\times 300$1 pixels, 2×2 analog islands, charge-sensitive preamplifiers, and AC-coupled 2-bit flash-type ADCs (Badea et al., 2024). Digital logic around the analog islands forms “featurized clusters,” notably the cluster $300\times 300$2-profile, position $300\times 300$3, and $300\times 300$4-size, and a neural classifier decides whether the cluster is compatible with a particle above a modest momentum threshold.

Three models are explored in simulation; the implemented choice is Model 2, a two-layer dense neural network that uses the full cluster $300\times 300$5-profile and reaches $300\times 300$6 acceptance for tracks with $300\times 300$7, compared with roughly $300\times 300$8 for the two-input Model 1 and an additional $300\times 300$9 signal-efficiency gain for the more complex CNN-based Model 3 (Badea et al., 2024). The resulting bandwidth saving for Model 2 is estimated to be $22.8$0–$22.8$1, the neural logic consumes about $22.8$2 per $22.8$3 array at an estimated occupancy of 1 hit/mm$22.8$4, and total power consumption is $22.8$5 per pixel, or approximately $22.8$6 (Badea et al., 2024). Initial tape-out testing validates correct timing through a 768-register chain up to 10 MHz, with 40 MHz remaining the design target (Badea et al., 2024). In this line of work, SmartDetector denotes a shift from passive sensing to selective, on-sensor data reduction.

6. Interpretable similarity detection for smart contracts

Outside physical sensing, SmartDetector denotes an interpretable function-similarity detector for smart contracts (Liu et al., 11 Sep 2025). The system targets clone detection, vulnerability propagation tracing, and fine-grained auditing at the Solidity function level. Its central idea is to decompose a function AST into statement trees rather than compare whole ASTs. Six statement-tree types are defined: Variable Definition, Assignment Operation, Conditional Block, Control Loop, Function Call, and Other Operation (Liu et al., 11 Sep 2025). For each tree, nodes are grouped into seven semantic categories—arithmetic operator, member variable, value, identifier, unit, data type, and code constructs—and category-level features are concatenated into a fixed-size representation (Liu et al., 11 Sep 2025).

Pairwise statement-tree similarity is then classified by LightGBM, but the classifier’s seven hyperparameters are optimized by a cosine-wise diffusion process rather than default settings, grid search, or Bayesian optimization. At the function level, if

$22.8$7

SmartDetector computes a score matrix $22.8$8 over all statement-tree pairs and aggregates it into

$22.8$9

declaring the functions similar if either score exceeds $30.1$0 (Liu et al., 11 Sep 2025). The paper reports an overall average F1-score of $30.1$1 and an average improvement of $30.1$2 in F1-score over current state-of-the-art methods, with $30.1$3 F1 on FC-pairs, $30.1$4 on BL-pairs, and $30.1$5 on cross-platform pairs from Binance Smart Chain and Polygon (Liu et al., 11 Sep 2025). Because statement trees are line-mappable, the system also highlights which lines in two functions are similar, making the output directly usable for auditors.

7. Common principles, misconceptions, and trajectory

A common misconception is that SmartDetector names a single standardized architecture. The literature instead uses it for multiple detector classes whose shared property is explicit accommodation of a domain bottleneck. Some systems remain fully model-driven and rule-based, such as the radar-only pipeline using thresholding, clustering, and rule-based 3D classification (Liu et al., 19 Jan 2026). Others are hybrid systems with simple heuristics plus learned components, such as the AGV-based environmental classifier paired with Backbone17-Det (Tran et al., 2021). Others push intelligence directly into hardware through on-chip neural inference and local data reduction (Badea et al., 2024). Still others make interpretability central, as in statement-level smart-contract clone analysis (Liu et al., 11 Sep 2025).

A second misconception is that “smart” necessarily means end-to-end deep learning. The hazard-sensing works show the opposite: triple-GEM UV imaging and multifunctional avalanche gaseous detectors are “smart” because they combine high sensitivity, localization, event discrimination, and networked operation, not because they rely on deep neural training (Volpe et al., 2017, Abbrescia et al., 2020). Likewise, the self-aware object detection framework distinguishes image-fidelity monitoring from ordinary detection confidence: the degradation score is defined geometrically in representation space and is explicitly independent of the detector’s box-confidence output (Becker et al., 20 Feb 2026).

Taken together, these works suggest a stable research trajectory. SmartDetector systems increasingly combine three layers: a task-specific perception core, a mechanism that models the dominant source of unreliability or inefficiency, and a reporting or control interface that exposes this extra intelligence to downstream users or infrastructure. In thermal imagery that extra layer is feature fusion for small targets (Zhang et al., 2021); in ADAS and mobile vision it is context monitoring or resource-aware control (Tran et al., 2021, Callegaro et al., 2022); in harsh industrial sensing it is modality choice and model-driven filtering (Liu et al., 19 Jan 2026); in front-end electronics it is on-chip selection under bandwidth constraints (Badea et al., 2024); and in program analysis it is statement-level explainability over structural similarity (Liu et al., 11 Sep 2025). A plausible implication is that future uses of the term will continue to emphasize not generic detection strength, but explicit handling of the failure mode that matters most in the target deployment.