Towards Practical Precision Agriculture: Real-Time Fruit Detection and Video Analytics on Embedded Edge Hardware

Ivica Dimitrovski; Vlatko Spasev; Ivan Kitanovski; Petre Lameski; Dane Boshev

cs.CVSep 11, 2026

Towards Practical Precision Agriculture: Real-Time Fruit Detection and Video Analytics on Embedded Edge Hardware

Authors: Ivica Dimitrovski, Vlatko Spasev, Ivan Kitanovski, Petre Lameski, Dane Boshev

Organizations: Faculty of Computer Science and Engineering, University Ss Cyril and Methodius, ul. Rudzer Boshkovikj 16, P.O. 393, Skopje, 100, North Macedonia · Faculty of Agricultural Sciences and Food, University Ss Cyril and Methodius, 16-ta Makedonska brigada br. 3, Skopje, 100, North Macedonia

Abstract

Static-image benchmarks do not capture the computational and temporal requirements of practical orchard video analytics. This study presents an end-to-end framework for real-time fruit detection, tracking, and counting on the NVIDIA Jetson Orin Nano Super. A lightweight YOLO26s detector is trained independently on four public datasets representing apples, mangoes, blueberries, and strawberries under a common protocol. The models are deployed on embedded platform using PyTorch and TensorRT at FP32, FP16, and INT8 precision. APPLE MOTS is then used for temporal video analytics because it provides orchard sequences with persistent fruit identities, enabling evaluation of multi-object tracking and unique-fruit counting. The selected FP16 TensorRT detector is integrated into an NVIDIA DeepStream pipeline combining hardware-accelerated decoding, ByteTrack tracking, and motion-aware line-crossing analytics. Across the four detection tasks, mean test mAP@50:95 ranges from 0.4957 to 0.8656. On the Jetson, TensorRT FP16 achieves 66.76-74.56 images/s at 13.41-14.98 ms prediction latency, while reducing mAP@50:95 by only 0.0020-0.0054 and gross energy consumption by approximately 64-66% relative to PyTorch FP32. The complete detector-tracker-analytics pipeline reaches 44.96-54.11 FPS and sustains the configured 30-FPS input rate without output-frame loss. On held-out orchard video sequences, HOTA ranges from 0.345 to 0.538, event-level counting F1 from 0.611 to 0.803, and relative count error from 6.2% to 51.6%. Performance varies across acquisition geometries: near-lateral row viewing yields the most stable tracking and counting, whereas forward traversal remains association- and recall-limited despite spatially adaptive counting geometry. These results show that practical edge-based fruit monitoring requires efficient detection and acquisition geometries that support reliable temporal association.

Figures & tables

Figure 1: Overview of the breadth-to-depth experimental workflow. A common YOLO26s detector is first evaluated across four fruit datasets, followed by reference-GPU characterization on an NVIDIA GeForce RTX 5060 Ti with 16 GB of graphics memory and embedded PyTorch/TensorRT optimization on an NVIDIA Jetson Orin Nano Super with 8 GB of unified memory. The selected FP16 TensorRT configuration is then integrated into a DeepStream-based video pipeline with ByteTrack association and calibration-derived motion-aware fruit counting.

Figure 2: Representative examples from the four fruit datasets: (a) DeepBlueberry, (b) APPLE MOTS, (c) StrawDI_Db1, and (d) MangoYOLO. The datasets exhibit substantial variation in fruit scale, object density, occlusion, native image resolution, background complexity, and acquisition viewpoint.

Dataset	Precision	Recall	mAP@50	mAP@50:95
MangoYOLO	$0.9642\pm 0.0045$	$0.9617\pm 0.0044$	$0.9904\pm 0.0004$	$0.7197\pm 0.0036$
DeepBlueberry	$0.7982\pm 0.0136$	$0.7684\pm 0.0042$	$0.8484\pm 0.0033$	$0.6813\pm 0.0066$
StrawDI_Db1	$0.9734\pm 0.0022$	$0.9357\pm 0.0039$	$0.9756\pm 0.0017$	$0.8656\pm 0.0017$
APPLE MOTS	$0.7692\pm 0.0174$	$0.7392\pm 0.0176$	$0.8189\pm 0.0232$	$0.4957\pm 0.0184$

Table 2: Test-set detection performance of YOLO26s across five independent training runs. Values are mean

\pm

sample standard deviation.

	PyTorch FP32 / FP16			TensorRT FP32 / FP16
Dataset	Inference (ms)	Prediction (ms)	E2E (ms)	Inference (ms)	Prediction (ms)	Throughput (img/s)	mAP @50:95	Energy (J/image)
MangoYOLO	5.908 / 6.112	7.222 / 7.425	13.117 / 13.068	2.235 / 1.048	3.527 / 2.340	283.57 / 427.43	0.7203 / 0.7207	0.2697 / 0.1603
DeepBlueberry	6.074 / 6.289	7.647 / 7.884	23.449 / 23.250	2.218 / 1.010	3.713 / 2.494	269.31 / 401.04	0.6705 / 0.6707	0.2537 / 0.0923
StrawDI_Db1	5.939 / 6.145	7.397 / 7.601	17.964 / 18.160	2.237 / 1.051	3.622 / 2.438	276.07 / 410.25	0.8591 / 0.8587	0.2719 / 0.1601
APPLE MOTS	5.941 / 6.135	7.464 / 7.763	25.466 / 25.200	2.214 / 1.049	3.717 / 2.550	269.01 / 392.09	0.4935 / 0.4924	0.3000 / 0.1363

Table 3: Reference-GPU runtime results on the NVIDIA GeForce RTX 5060 Ti. Paired values denote FP32/FP16 execution. Inference latency denotes model execution, while prediction latency measures the CUDA-synchronized preprocessing-to-detections path for an image already loaded in memory. PyTorch end-to-end (E2E) latency additionally includes file reading and image decoding. TensorRT throughput is derived from prediction latency. Energy is derived from GPU-board power and should therefore not be interpreted as directly comparable to the whole-module energy measurements reported for the Jetson platform.

Dataset	Configuration	mAP @50:95	mAP change	Inference (ms)	Prediction (ms)	Throughput (img/s)	Power (W)	Energy (J/image)
MangoYOLO	PyTorch FP32	0.7228	$0.0000$	29.244	36.671	27.27	13.06	0.4789
	PyTorch FP16	0.7233	$+0.0005$	29.511	36.795	27.18	10.73	0.3948
	TensorRT FP32	0.7205	$-0.0023$	12.151	19.464	51.38	14.73	0.2867
	TensorRT FP16	0.7208	$-0.0020$	6.157	13.412	74.56	12.30	0.1650
	TensorRT INT8	0.7015	$-0.0213$	6.128	13.415	74.54	10.63	0.1427
DeepBlueberry	PyTorch FP32	0.6730	$0.0000$	29.113	37.215	26.87	12.91	0.4805

Table 4: YOLO26s deployment results on the NVIDIA Jetson Orin Nano Super. The mAP change is reported relative to PyTorch FP32 for the corresponding validation-selected checkpoint. Inference latency denotes model execution, while prediction latency measures the CUDA-synchronized preprocessing-to-detections path for an image already loaded in memory, including inference and postprocessing but excluding file reading and image decoding. Throughput is derived from prediction latency. Power and energy correspond to whole-module Jetson measurements.

Figure 3: Representative acquisition geometries for the held-out APPLE MOTS video sequences, shown from left to right as 0006, 0007, and 0008. Sequence 0006 is acquired during forward traversal between two orchard rows and is therefore divided into two independent spatial motion regions; sequence 0007 contains an oblique reverse traversal; and sequence 0008 provides an approximately lateral view of a single orchard row. Blue arrows indicate the dominant image-plane motion estimated from ByteTrack calibration trajectories, rather than the physical camera-motion vector. Green finite lines indicate the corresponding motion-aware counting gates. The dashed partition in sequence 0006 separates the two independently modeled spatial regions.

Sequence	MOTA	IDF1	HOTA	DetA	AssA	IDSW	Frag.
0006	0.081	0.477	0.345	0.292	0.411	109	146
0007	0.442	0.557	0.378	0.362	0.397	144	133
0008	0.363	0.665	0.538	0.427	0.681	34	109

Table 5: Multi-object tracking performance on the held-out APPLE MOTS sequences using the validation-selected ByteTrack configuration. HOTA, DetA, and AssA denote the standard Higher Order Tracking Accuracy decomposition. IDSW denotes identity switches and Frag. denotes trajectory fragmentations.

Seq.	Gates	GT	Pred.	TP	FP	FN	Precision	Recall	F1	Count error
0006	2	64	31	29	2	35	0.935	0.453	0.611	51.6%
0007	1	68	42	42	0	26	1.000	0.618	0.764	38.2%
0008	1	81	76	63	13	18	0.829	0.778	0.803	6.2%
Overall	–	213	149	134	15	79	0.899	0.629	0.740	30.0%

Table 6: Event-level fruit-counting performance on the held-out APPLE MOTS sequences. GT and Pred. denote reference and predicted crossing counts. TP, FP, and FN are obtained by one-to-one spatiotemporal event matching. Count error is the absolute difference between the predicted and reference crossing counts, normalized by the reference crossing count.

Seq.	Throughput	PGIE	ByteTrack	Analytics	Dec. $\rightarrow$ Track	Dec. $\rightarrow$ Analytics	Power	Energy/frame
	(FPS)	(ms)	(ms)	(ms)	(ms)	(ms)	(W)	(J)
0006	44.96	18.83	18.99	1.67	151.60	153.27	8.41	0.187
0007	47.88	17.52	19.26	2.14	156.14	158.28	9.01	0.188
0008	54.11	15.97	14.59	1.41	121.03	122.44	10.14	0.187

Table 7: Maximum-throughput end-to-end video-processing performance on the NVIDIA Jetson Orin Nano Super. PGIE denotes the TensorRT detector stage. Decoder-to-tracker and decoder-to-analytics values represent frame residence latency through the pipelined processing graph and should not be interpreted as inverse throughput. PGIE timing represents the detector stage within the running DeepStream pipeline and is therefore not directly equivalent to the standalone TensorRT inference latency reported in Table 4 .

Explore similar work

Apr 11, 2026

A Comparative Study of Modern Object Detectors for Robust Apple Detection in Orchard Imagery

Accurate apple detection in orchard images is important for yield prediction, fruit counting, robotic harvesting, and crop monitoring. However, changing illumination, leaf clutter, dense fruit clusters, and partial occlusion make detection difficult. To provide a fair and reproducible comparison, this study establishes a controlled benchmark for single-class apple detection on the public AppleBBCH81 dataset using one deterministic train, validation, and test split and a unified evaluation protocol across six representative detectors: YOLOv10n, YOLO11n, RT-DETR-L, Faster R-CNN (ResNet50-FPN), FCOS (ResNet50-FPN), and SSDLite320 (MobileNetV3-Large). Performance is evaluated primarily using COCO-style [email protected] and [email protected]:0.95, and threshold-dependent behavior is further analyzed using precision-recall curves and fixed-threshold precision, recall, and F1-score at IoU = 0.5. On the validation split, YOLO11n achieves the best strict localization performance with [email protected]:0.95 = 0.6065 and [email protected] = 0.9620, followed closely by RT-DETR-L and YOLOv10n. At a fixed operating point with confidence >= 0.05, YOLOv10n attains the highest F1-score, whereas RT-DETR-L achieves very high recall but low precision because of many false positives at low confidence. These findings show that detector selection for orchard deployment should be guided not only by localization-aware accuracy but also by threshold robustness and the requirements of the downstream task.

Mohammed Asad, Ajai Kumar Gautam, Priyanshu Dhiman +1

Dept. of Electronics and Communications Engineering, Delhi Technological University, New Delhi, India

Object Detection Agricultural Image Analysis

Ask PDF

Feb 22, 2026

A Two-Stage Detection-Tracking Framework for Stable Apple Quality Inspection in Dense Conveyor-Belt Environments

Industrial fruit inspection systems must operate reliably under dense multi-object interactions and continuous motion, yet most existing works evaluate detection or classification at the image level without ensuring temporal stability in video streams. We present a two-stage detection-tracking framework for stable multi-apple quality inspection in conveyor-belt environments. An orchard-trained YOLOv8 model performs apple localization, followed by ByteTrack multi-object tracking to maintain persistent identities. A ResNet18 defect classifier, fine-tuned on a healthy-defective fruit dataset, is applied to cropped apple regions. Track-level aggregation is introduced to enforce temporal consistency and reduce prediction oscillation across frames. We define video-level industrial metrics such as track-level defect ratio and temporal consistency to evaluate system robustness under realistic processing conditions. Results demonstrate improved stability compared to frame-wise inference, suggesting that integrating tracking is essential for practical automated fruit grading systems.

Keonvin Park, Aditya Pal, Jin Hong Mok

Industrial Inspection Video Object Detection

Ask PDF

Aug 23, 2026

Optimizing YOLO27, YOLO26, YOLO11, and YOLOv8 for Fine-Grained Small-Object Detection and Segmentation in Complex Orchard Environments

This study presents an architectural and experimental cross-generation analysis of Ultralytics YOLO27 (YOLOv27), YOLO26 (YOLOv26), YOLO11 (YOLOv11), and YOLOv8 for fine-grained robotic perception in complex orchard environments. Fine-grained detection and instance segmentation of early-stage fruit anatomy remain challenging in complex orchard environments because of limited pixel footprints, green-on-green similarity, occlusion, and substantial scale variation. Because YOLO27 has been announced but its public implementation and trainable segmentation models are not yet available, the present study provides an architectural analysis of YOLO27, while controlled experiments benchmark YOLOv8, YOLO11, and YOLO26; YOLO27 experiments will be incorporated following public model availability. Five model scales-nano (n), small (s), medium (m), large (l), and extra-large (x)-were evaluated for fruitlet, calyx, and peduncle detection and segmentation using conventional 640 x 640 and small-object-focused 960 x 960 configurations, yielding 30 experiments. YOLO11s-960 achieved the highest observed mask mAP@50:95(0.402) and box mAP@50:95(0.426), whereas YOLO26s-960 achieved comparable values of 0.397 and 0.425 with only 10.37~~M parameters and 34.1~~GFLOPs. Peduncle remained the most challenging class, and increasing model capacity did not consistently improve accuracy. Overall, compact-to-moderate YOLO models combined with small-object-focused training provided favorable accuracy-efficiency trade-offs, establishing a reproducible benchmark for fine-grained agricultural robotic perception. Code, trained models, and experimental configurations are publicly available, and will be updated through our Github Link: https://github.com/rnjnspkt/Optimizing-and-Comparing-Ultralytics-YOLOv26-YOLOv11-and-YOLOv8-for-Small-Object-Detection-and-Seg

Ranjan Sapkota, Manoj Karkee

Cornell University, AgRobotics Lab, 134 Riley-Robb Hall, Ithaca, NY, 14853, USA

Small Object Detection Instance Segmentation Fine-Grained Visual Perception You Only Look Once Agricultural Robotics Robotic Perception

Ask PDF