Due to linear perspective effects, the acquired video data in this study exhibits distortion, with rapeseed seedling density appearing denser in the distance and sparser in the foreground along the shooting direction, and the pixel values occupied by rapeseed seedlings showing smaller in the distance and larger in the foreground, as shown in Figure 5a.

DBSCAN is a classic density-based unsupervised clustering algorithm that enables data-driven clustering via neighborhood density thresholds (i.e., clustering radius eps and minimum sample count min_samples) (Ester et al., 1996; Cheng et al., 2024). However, the two core parameters of the original DBSCAN algorithm are fixed and necessitate manual tuning, which renders it inadequate to accommodate the “small in distance and large in foreground” perspective distortion of rapeseed seedlings in oblique-view videos, resulting in erroneous clustering in areas with adjacent density variations. Existing adaptive DBSCAN variants (e.g., Khan et al., 2018) are predominantly optimized for general-purpose scenarios and do not account for the unique density distortion characteristics of crops under field oblique viewing conditions. To address the perspective-induced “small in distance and large in foreground” challenge, an adaptive DBSCAN parameter-tuning method (AdapDBSCAN) is proposed. A parameter optimization method based on local density was introduced in this study, constructing a nonlinear mapping relationship between the eps parameter and regional target density, while establishing a dynamic calculation model for the min_samples parameter. The center point of the detection box is taken as the position of each detected rapeseed seedling, with each center point treated as a sample point representing a rapeseed seedling. Firstly, a nonlinear functional relationship between the eps parameter and local regional target density is constructed to dynamically adjust the eps: when the number of samples in a region is dense, the eps value decreases; when the number of samples is sparse, eps increases. Secondly, the number of samples within the eps is counted to dynamically calculate the min_samples parameter: when the number of samples in a region is denser, the number of core point min_samples is larger; when the number of samples is sparser, the number of core point min_samples is smaller. Finally, based on this functional relationship, starting from any point x, if the number of samples within the eps of point x is greater than or equal to min_samples, point x is marked as a core point; otherwise, it is marked as a noise point. Core points form clusters, and noise points are marked as candidate elimination set N and eliminated during clustering, as shown in Figure 5b. Through dynamic determination of parameters twice, rapeseed seedlings in non-target regions (i.e., sparse seedlings detected on both sides) are eliminated. Specific formulas are shown in Equations 1, 2. The process is shown in Figure 5c, with the red area representing the target region.

where Δv is the vertical span, α is a normalization constant utilized to map the vertical span Δv to a scaling factor; median(D_filtered) is the filtered horizontal spacing of rapeseed seedlings (i.e., filtering out outliers such as excessively large spacing) to avoid interference of extreme values on the reference distance. ρ_x is the neighborhood density of the x sample within the eps, and δ is the density decay factor.

In addition, rapeseed seedlings in non-target regions have small imaging sizes and sparse distribution in images. If not controlled, DBSCAN is forced to continuously increase eps until covering the entire image to meet the min_sample threshold, leading to clustering failure (all points are classified into the same cluster or noise). In this study, the image size is 1920×1080 pixels. Considering that metric coordinates in images are depth-dependent, a global pixel-to-actual-distance mapping was not adopted. Instead, based on camera calibration at a fixed shooting height (0.5 m) and actual measurement of the target region, the average pixel-to-distance ratio in the central area of the image (where target seedlings are concentrated) was approximately 800 pixels per meter. When the maximum threshold of eps exceeds the width of the drainage ditch, a broader non-target area enters. Here, the spacing between the rapeseed seedlings has further increased. Even if the eps is further expanded, it is difficult to gather min_sample samples within the neighborhood. Eventually, it falls into a cycle of infinitely increasing the eps to meet the threshold. Based on this, according to the image spatial scale and the actual width of the drainage ditch (0.3 meters), the maximum threshold of eps was preset to 250 pixels. The corresponding actual space is approximately 0.3125 meters, which has exceeded the width of the drainage ditch. And a dual-threshold collaborative constraint mechanism was constructed with the filtered horizontal spacing of rapeseed seedlings (median(D_filtered)) to effectively suppress unlimited expansion of the eps parameter.

Representative algorithms were compared via controlled experiments. The SORT algorithm was ultimately selected.

Integrating previous research, the rapeseed seedling counting method integrated YOLOv11, AdapDBSCAN, and SORT to realize the counting of rapeseed seedlings in target regions. To better count rapeseed seedlings, when the geometric center of a counting box (for counting seedlings) crosses the lower edge of the video frame, the counting box is automatically removed, and its unique identifier (ID) assigned by the SORT algorithm is permanently marked to ensure no reuse in subsequent frame processing, thus ensuring ID uniqueness (Wu et al., 2023b). This counting method eliminates non-target rows through the AdapDBSCAN algorithm, only counts target regions, and displays the current number of detected rapeseed seedlings and frame number in the upper left corner of the video.

The mobile client implementation of the CropTriangulator system adopts a three-tier architectural framework comprising: (1) a user interface (UI) layer built with native UI components for video upload and visualization; (2) an interaction layer managing file selection and processing requests through native event handlers; and (3) a presentation layer generating dynamic result displays. The system incorporates mobile-optimized UI design with screen-adaptive layouts, file picker functionality providing visual feedback, and animated processing indicators. The user operation interfaces are illustrated in Figure 6.

The operational workflow begins when users upload field-captured videos through the functional interface. These video files are transmitted to the backend server via HTTP protocol, where they are queued for processing. The server executes the complete CropTriangulator analysis pipeline, including YOLOv11-based seedling detection, AdapDBSCAN clustering for target region identification, and SORT algorithm for seedling tracking and counting. Upon completion, the processed results are returned to the client application. Users can view the rapeseed seedling count calculated by the CropTriangulator method on the counting results interface, as shown in Figure 6c. The entire client application is implemented with native mobile development technologies to ensure platform compatibility and maintainability.

Different scales of YOLOv11 models all performed well in rapeseed seedling detection, as they balanced AP and inference time. The AP and inference time of different scales of YOLOv11 models from different angles are summarized in Table 3, with the best scores highlighted in bold. As the number of parameters in the feature extraction network increased, AP_0.5 and AP_0.5:0.95 for both 45° and 90° views showed a downward trend, and inference time also increased. The nano-scale and medium-scale models for the 45° view had the same AP_0.5 and AP_0.5:0.95 values, but the medium-scale model had an inference time 115% slower than the nano-scale model, while the nano-scale model achieved the fastest inference time and required the smallest model size. For the 90° view, the nano-scale model had the highest AP_0.5 and AP_0.5:0.95 and the shortest inference time. Figure 7 shows visualization heatmaps generated by Grad-CAM (Yang Y. et al., 2025), intuitively presenting the regions focused on by target detection models of different scales. The brightness of each region indicates its importance to the model, with higher brightness representing stronger attention. Results showed that the nano-scale model performed best for the 45° view, and the small-scale model performed best for the 90° view, but with little difference from the nano-scale model. YOLOv11n outperforms larger-scale YOLOv11 models due to its superior adaptation to late-stage rapeseed seedling detection scenarios, rather than being constrained by dataset scale. Late-stage rapeseed seedlings exhibit relatively uniform morphological and spectral features compared to general targets, leading larger models with excessive parameters to overfit trivial background noise (e.g., soil texture variations, isolated weed pixels) in field images. In contrast, the streamlined architecture of YOLOv11n (incorporating Focus and C3k2 modules) prioritizes the extraction of core seedling features (e.g., leaf contours, stem-root connections), enabling it to achieve the same AP_0.5 and AP_0.5:0.95 values as the medium-scale model. By comprehensively considering detection accuracy, computational efficiency, and heatmap results, the nano-scale model has advantages in detecting small-scale targets, so it was selected for target detection in both 45° and 90° views.

However, the superior performance of YOLOv11 is affected to some extent, leading to detection boxes in unannotated regions. For example, when the model scans complex scenes, its strong feature extraction capability captures elements in unannotated regions that are similar to target features, as shown in Figure 8a. This not only generates false detection results in unmarked regions but also complicates subsequent result screening, as shown in Figure 8b. Additionally, YOLOv11’s excellent noise reduction and detail enhancement capabilities may parse the fuzzy pixel clusters in unannotated regions into recognizable target forms, causing detection boxes in non-target regions, which interfere with the detection accuracy of the model.

Although some studies have improved model detection performance using multi-source data such as depth information and infrared imaging (Stumpe et al., 2025; Yang Y. et al., 2025), most current smartphones (especially mid-to-low-end models accounting for the majority of the market) do not have dedicated hardware depth sensors (e.g., ToF, LiDAR). While depth can be inferred from stereo cameras based on multi-view geometry, this method is constrained by environmental interference, high computational complexity, and poor adaptability to dynamic field shooting. By focusing on smartphone RGB cameras, convenient data collection has ultimately been achieved for users; as more users upload RGB video data to the cloud, the dataset can be continuously enriched to boost model generalization. To ensure the universality, real-time performance, and scalability of the method, the counting pipeline has been optimized based solely on RGB video. Therefore, with the continuous improvement of future phone performance, smartphones are expected to break through existing hardware limitations and become terminal devices capable of directly acquiring, processing multi-source data, and outputting results.

Among different scales of YOLOv11 models, target detection results from the 45° view were significantly better than those from the 90° view, with P, AP_0.5, and AP_0.5:0.95 for the 45° view significantly higher than those for the 90° view. In particular, for AP_0.5:0.95, the selected nano-scale model for the 45° view improved by more than 30% compared to the 90° view, with little difference in response time, providing guidance for tracking and counting performance from different views. As shown in Table 3, the precision of the nano target detection model selected for the two oblique views in rapeseed scenarios was 90.8% and 77.5%, respectively, with recall rates of 85% and 80.2%. High precision and stable recall rates provide a basis for reducing target loss during tracking and counting, avoiding overestimation of counting results due to misjudgment of targets.

The AdapDBSCAN clustering method (adaptive method) based on SORT and DeepSORT exhibited excellent counting performance in processing rapeseed seedling videos captured by smartphones. In addition, the counting method using fixed-parameter DBSCAN clustering based on SORT and DeepSORT (conventional method) showed considerable drawbacks. As shown in Figure 9, counting results using the adaptive method for Videos 1 to 20 were closer to the GT values, while the conventional method had larger deviations compared to the adaptive method. Taking Video 4 (GT = 680) as an example, the counting result using the CropTriangulator method was 679, with an accuracy of 99.85%, and the result using the AdapDBSCAN-DeepSORT method was 554, with an accuracy of 81.47%. Compared to the DBSCAN-SORT and DBSCAN-DeepSORT methods, the accuracy increased by 3.53% and 26.32%, respectively. A comparison of randomly selected frames from Video 4 using the DBSCAN-SORT method and the CropTriangulator method is shown in Figure 10. The adaptive method effectively filtered out detected rapeseed seedlings in non-target rows, improving counting accuracy. These results demonstrate the excellent performance and feasibility of the adaptive method.

In a certain frame of the video, due to the geometric characteristics of perspective projection, the actual spatial distribution of distant rapeseed seedlings is compressed in the image, which is manifested as an increase in the number of seedlings per unit image area (i.e., visually increasing density); at the same time, distant rapeseed seedlings occupy fewer pixels in the image, reflecting the impact of projection scaling on target imaging size (Wang Z. et al., 2022). In contrast, due to the short distance from the camera, foreground rapeseed seedlings are less affected by perspective projection scaling, showing sparse spatial distribution in the image, fewer seedlings per unit image area, and more pixels occupied by individual seedlings. This imaging feature of “denser in distance and sparser in foreground” leads to uneven density distribution of acquired rapeseed seedling data (Islam et al., 2024). The conventional method cannot adapt to local density differences, resulting in incorrect clustering and failure to filter out detection results from unmarked regions in the target detection stage, which are instead classified as target row rapeseed seedlings, leading to lower accuracy of the conventional method compared to the adaptive method.

The proposed adaptive method (AdapDBSCAN) fully leverages local target density information and a dual-threshold constraint mechanism to address these inherent limitations: it constructs a nonlinear mapping between the clustering radius (eps) and regional seedling density, dynamically shrinking eps in dense distant regions to maintain precise clustering and expanding eps in sparse foreground regions to preserve valid seedlings, while synchronously adjusting the min_samples parameter based on local sample counts. This scenario-specific optimization enables the adaptive method to accurately identify and separate target rows from non-target areas (e.g., drainage ditches and adjacent rows) even under perspective-induced density variations, effectively filtering out misdetected seedlings in unmarked regions that the conventional method fails to exclude. This approach significantly reduces erroneous clustering associated with fixed parameters, enhances the precision of non-target seedling elimination on both sides of the target rows, and ultimately achieves higher counting accuracy and stability compared to traditional approaches, fully demonstrating its superiority in handling the uneven density distribution of rapeseed seedlings in oblique-view videos.

Perspective distortion introduces three key challenges in practical applications: in the 45° view, the foreground (0-2m from the camera) has a seedling density of 2–3 plants per m² (sparse), while the background (8-14m) reaches 8–10 plants per m² (dense), and fixed-parameter clustering (e.g., traditional DBSCAN) either over-clusters dense regions (merging adjacent seedlings) or under-clusters sparse regions (classifying valid seedlings as noise), leading to ±15% counting error; distant seedlings (background) occupy 3–8 pixels, while foreground seedlings occupy 30–50 pixels, and YOLOv11n’s detection confidence for small targets (less than 10 pixels) drops by 35%, increasing FN; the transition zone between target rows and drainage ditches (depth 12-14m) suffers from blurred edges due to perspective compression, leading to 10-12% of FP from misclassifying ditch edges as seedlings.

By applying AdapDBSCAN for algorithm optimization, these problems can be solved. The dual-threshold constraint (eps_max=250 pixels and median₍D_filtered)) dynamically adjusts clustering parameters based on local density. For dense background regions, eps is reduced to 50–80 pixels to avoid over-clustering, and for sparse foreground regions, eps is increased to 150–200 pixels to retain valid seedlings which reduces perspective-induced error by 42% compared to fixed-parameter DBSCAN.

Furthermore, to compare counting accuracy from different oblique angles, this study systematically evaluated performance differences between 45° oblique view and 90° vertical view. Given the significant advantages of target detection from the 45° view over the 90° view in performance indicators, it is reasonable to infer that superior performance is also expected in target tracking and final result output. Detection from the 45° view can more accurately capture target features, reducing error accumulation in key links such as feature matching and motion prediction, ultimately improving overall tracking performance. Experimental results are shown in Figure 9, indicating that for the 45° view, the average accuracy of the DBSCAN-SORT method (94.44%) was approximately 19% higher than that for the 90° view (75.59%); the average accuracy of the CropTriangulator method (97.13%) was more than 14% higher than that for the 90° view (82.94%). Additionally, as shown in Figure 9, for the 45° oblique view using the CropTriangulator method, 5 out of 10 videos had an accuracy of over 98%, with the highest at 99.85%, closely matching manual counting results and far exceeding the counting accuracy of the 90° vertical view using the CropTriangulator method. Moreover, linear regression analysis was applied between counting results obtained using the CropTriangulator method and GT. As shown in Figure 11, the R² of counting results from the 45° oblique view was 0.917. These results highlight the excellent counting performance of the 45° oblique view and its strong stability, fully verifying that the 45° oblique view is superior for counting tasks.

Differences in counting performance between angles may stem from three key factors: firstly, the overlap rate of seedlings increases from the vertical view, with foreground seedlings occluding background seedlings; secondly, perspective distortion is enhanced from the 90° vertical view, increasing distortion; finally, rapeseed seedlings from the 45° oblique view contain both root and leaf features, enabling more efficient handling of complex scene interference and accurate counting by the SORT algorithm, while the 90° vertical view contains highly similar crop root features, causing false detections that severely affect target detection and counting performance (Kumar et al., 2014; Liu et al., 2024).

Therefore, future research should focus on constructing more complex datasets containing more detailed and variable video acquisition views. In the future, we will adopt an image distortion correction model to correct the position of each pixel in the image through geometric transformation, restoring spatial relationships between pixels and improving data density consistency (Li et al., 2019). Additionally, efforts should be made to improve the efficiency and flexibility of counting methods, especially in environments with uneven density distribution.

Having established the superiority of the adaptive method and the 45° viewing angle for the framework, the subsequent step involved selecting the most suitable tracking component. For this purpose, control experiments were conducted on various tracking algorithms under the same experimental conditions. Experimental results indicated that BotTrack achieved an average W_ID of 42.56% and a counting accuracy of 59.70% for the 45° view, while ByteTrack exhibited a W_ID of 28.35% and an accuracy of 80.25% for the same view. Both algorithms showed higher W_ID and lower counting accuracy compared to SORT. The W_ID of ByteTrack, although lower than that of DeepSORT, remained substantially higher than the benchmark of 8.47% set by SORT, as shown in Table 4. The W_ID (8.47%) of SORT was 19.88% lower than that of ByteTrack and 34.10% lower than that of BotTrack. Furthermore, its counting accuracy (97.13%) was 16.88% and 37.43% higher than those of ByteTrack and BotTrack, respectively.

The inferior tracking stability of BotTrack and ByteTrack in this study can be attributed to two scenario-specific factors. First, despite its plant-targeted design, the feature matching module in BotTrack struggles to distinguish late-stage rapeseed seedlings with highly uniform morphology, leading to frequent ID misassignments and the highest W_ID (42.56%) among all tested algorithms. Its P_tr (61.30%) and P_mt (47.88%) are also the lowest, reflecting poor target association and matching performance under high-similarity conditions. Second, the strategy employed by ByteTrack for associating low-confidence detections, while beneficial in general crowded scenes, becomes counterproductive here. Since the preprocessing stage has already filtered out most noise, the remaining low-confidence boxes largely correspond to ambiguous patches or background clutter. Attempting to associate them introduces unnecessary computational overhead and increases the risk of ID switches, as reflected in its elevated W_ID (28.35%). Moreover, their inference speeds (BotTrack: 12.4 ms per frame; ByteTrack: 8.7 ms per frame) remain significantly slower than that of SORT (4.6 ms per frame), a critical disadvantage for real-time processing in practical deployment.

To focus on balancing efficiency and stability in practical applications and to avoid increasing computational burden and the risk of ID confusion due to the introduction of overly complex feature matching mechanisms, the subsequent comparison of tracking algorithms in this study was concentrated on SORT and DeepSORT. Although algorithms such as BotTrack and ByteTrack excel in general object tracking, their complex feature matching mechanisms were primarily designed for targets with significant appearance differences or variable motion patterns. Under the specific conditions where seedling features were nearly identical and motion was solely induced by camera movement, these sophisticated mechanisms not only struggled to provide effective discrimination but also could increase the risk of ID switches and computational overhead by relying on easily confusable appearance features. This approach aimed to more clearly demonstrate the advantages of a streamlined and efficient tracking framework (SORT) over deep trackers dependent on appearance features (DeepSORT) in the high-similarity scenario of rapeseed seedlings.

The SORT tracking algorithm exhibited good tracking performance in rapeseed fields. As shown in Figure 12, for 20 videos using the adaptive counting method based on SORT, the average ID switch rate (W_ID) was 8.47%, meaning most rapeseed seedlings (even those with severe mutual occlusion) maintained their IDs from appearance to disappearance, as shown in Figures 13a, c, e. In contrast, for the same videos with three consecutive frames using the adaptive counting method based on DeepSORT, IDs of rapeseed seedlings in dense regions changed significantly, as shown in Figures 13b, d, f. For the counting method based on SORT, the average target tracking accuracy (P_tr) and average target tracking precision (P_mt) were 87.53% and 89.75%, respectively. For videos using the counting method based on DeepSORT, the W_ID, P_tr, and P_mt were 36.05%, 71.50%, and 54.72%, respectively. The P_mt of the SORT-based counting method was 35.03% higher than that of DeepSORT, indicating that SORT is far superior to DeepSORT in terms of trajectory continuity and stability. Additionally, two consecutive frames randomly selected from a video using the DeepSORT tracking algorithm are shown in Figure 14. As can be seen from Figures 13b, d and 14, DeepSORT exhibits large fluctuations in tracking stability, which possibly due to its poor robustness in dense or occluded scenarios in this study.

Compared to the DeepSORT-based counting method, the SORT-based counting method was more accurate. As shown in Figure 9, counting results of the conventional and adaptive methods based on DeepSORT for Videos 1 to 10 were 0.55 and 0.79 times the GT value, respectively, while those for Videos 11 to 20 were only 0.4 and 0.45 times the GT value, far lower than the corresponding results of the SORT-based counting method. This may be because the DeepSORT-based counting method introduces a deep appearance feature extractor (Zhang et al., 2024), which struggles to learn subtle features that sufficiently distinguish highly similar plants. In the same video, IDs already assigned to rapeseed seedlings may be reassigned to new rapeseed seedlings, as shown in Figure 15, leading to generally lower counting results compared to the GT value. The highly similar appearance of rapeseed seedlings significantly affects counting accuracy, which is also the reason for the high W_ID of DeepSORT. Additionally, DeepSORT relies more on static background assumptions (Huang et al., 2024), while data in this study was acquired by staff walking in the field with a gimbal, with fixed rapeseed seedlings, undermining the reliability of DeepSORT’s motion prediction and appearance matching. In the environment of this study, the tracking performance of DeepSORT was far inferior to that of SORT. Therefore, the CropTriangulator method combining the adaptive method with SORT provides considerable advantages for real-time counting of rapeseed seedlings.

The main sources of counting errors in this study include target overlap, false positives (FP), and false negatives (FN): late-stage rapeseed seedlings have dense foliage, leading to 30-40% overlap in the 90° vertical view but 15-20% in the 45° oblique view, which causes the model to misclassify multiple seedlings as a single target and results in FN, for example, in Video 18 (90° view), 12 out of 38 FN errors (31.6%) were attributed to severe leaf overlap, while the 45° oblique view reduces overlap by exposing more 3D structural features (e.g., stem-root separation), and the SORT algorithm’s motion prediction further mitigates this by maintaining ID continuity for partially occluded seedlings; FP primarily originate from non-target regions (e.g., weeds, soil clods) and perspective-induced misdetection, and the AdapDBSCAN algorithm filters 89.2% of FP by dynamic density clustering, but residual FP (accounting for 2.3% of total counts) still occur in edge regions of drainage ditches due to the pixel-to-distance ratio in edge regions deviating from the central area, leading to over-clustering; beyond overlap, FN are caused by motion blur (18.7% of FN) and small-sized seedlings in distant regions (22.1% of FN). Motion blur (resulting from walking speed fluctuations) reduces detection confidence, while distant seedlings (less than 5 pixels in diameter) are easily missed by YOLOv11n.

To verify the superiority of the CropTriangulator method in late-stage rapeseed seedling counting, four representative SOTA crop counting methods were selected for benchmarking. The comparison is based on publicly reported performance metrics from the original studies, with a focus on counting accuracy. The comparative results were summarized in Table 5, where the performance of existing methods was extracted from their original publications, and the performance of CropTriangulator was based on the 45° optimal view.

As shown in Table 5, CropTriangulator method demonstrates significant advantages in late-stage rapeseed counting. Compared to the only rapeseed-specific method (Zhao et al., 2018), CropTriangulator achieves a 13.46% higher accuracy, solving the bottlenecks of single-image data (partial target missing) and late-stage occlusion that limit existing rapeseed counting methods. Despite targeting more complex late-stage rapeseed (with severe leaf overlap and perspective distortion), the accuracy of CropTriangulator is comparable to early-stage peanut counting (Lin et al., 2022) and outperforms dense cotton seedling counting (Tan et al., 2023), verifying strong adaptability to complex growth stages. Unlike Rong et al. (2023) which relies on depth sensors and controlled greenhouse environments, CropTriangulator achieves comparable accuracy in open fields.

These results confirm that the CropTriangulator method effectively addresses the optimization gap of existing methods in late-stage rapeseed seedling counting, balancing high accuracy, adaptability to complex field scenarios, and practical applicability.

Although the problem of data density distribution in rapeseed seedling counting has been basically solved, several challenges remain: firstly, motion blur and shooting area offset not only affect the detection stage but also impact clustering and counting performance. Secondly, the CropTriangulator method is constrained by tracking results. Some studies indicate that existing algorithms have poor tracking accuracy for dense pedestrians, suggesting poor performance of current tracking methods in complex and variable backgrounds (Wang W. et al., 2022). Fortunately, improvements in tracking algorithms have shown positive effects in kiwifruit orchard counting (Zhang et al., 2025). However, direct cross-study comparisons are challenging due to differences in data acquisition standards. Stable relative motion during data acquisition ensures effective cross-frame matching. It should be noted that rapeseed seedlings have similar phenotypic characteristics, especially those in the same growth stage, making it difficult for existing tracking technologies with feature recognition modules to distinguish them, leading to ID assignment errors. Some studies have achieved accurate crop identification by acquiring images of crops at different growth stages in the field (Li et al., 2025; Naseer et al., 2025).

Challenges in crop counting in complex field environments require comprehensive research in the future. The adaptive dynamic parameter adjustment mechanism of AdapDBSCAN must be optimized to improve clustering performance, especially in uneven density scenarios. User-friendly video acquisition methods should also be explored to better control video quality, reduce the impact of motion errors on clustering, and thus weaken their impact on counting. Meanwhile, developing stable and efficient rapeseed seedling tracking technologies in complex agricultural scenarios remains a topic worthy of further exploration. Additionally, future research is advised to focus on temporal analysis (lv et al., 2025; Wang et al., 2024), tracking morphological changes of the same plant at different seedling stages (early and late seedling stages), which may help improve the stability of tracking methods and avoid duplicate counting.

The video-based rapeseed seedling counting method proposed in this study is more convenient than previous image-based algorithms and more easily applicable to actual crop planting. These results verify the excellent performance of the CropTriangulator method and the feasibility of eliminating rapeseed seedlings in non-target regions based on target density distribution. Therefore, this study contributes to formulating strategies to improve crop emergence rates and productivity estimation.