Herbicide-free operation is often required in precision agriculture. A robotic laser weeding can also be expected for the precise object detection and stable keypoint localization for the weed growth points. The YOLOv8-Pose model has been applied to the visual detection of the laser weeding, particularly for the rapid detection and keypoint localization. However, the technical challenges still remained on the deployment in the complex field environments, for instance, the high degree of morphological similarity between crop seedlings and weeds during early growth stages, the substantial diversity in the weed target scales from the minute sprouts to mature plants, and the computational constraints to edge devices on mobile robotic platforms. The standard YOLOv8-Pose architecture cannot balance the requirements between the multi-scale feature extraction and real-time inference speed. In this study, a lightweight neural network, LSPKI-YOLO, was proposed and then validated for the visual perception and precise positioning in dynamic agricultural environments. Two architectures were integrated into the YOLOv8-Pose baseline in order to enhance the feature representation with less complexity. Firstly, the capacity was significantly improved to discern the subtle textural discrepancies between the crops and weeds of varying sizes. The standard feature extraction module was also replaced by a C2f-PKI module. A Poly Kernel Inception (PKI) block was synergistically fused with a Context Anchor Attention (CAA) mechanism. A series of parallel depth-wise separable convolutions was equipped with varying kernel sizes—ranging from the small to the large receptive fields—to extract the texture features at the multiple granularities, particularly without compromising a significant increase in the parameter count. Concurrently, the CAA mechanism employed the orthogonal strip convolutions to capture the long-range contextual dependencies, in order to effectively suppress the background noise caused by environmental interference, such as the leaf occlusion. Secondly, a Lightweight Shared-Convolution Batch Normalization (LSBN-Pose) detection head was designed to resolve the conflict between the model size and detection precision. A convolution weight mechanism was shared over the different feature scales for the model compression, unlike the conventional heads with independent branches for the different feature scales. Simultaneously, the independent Batch Normalization (BN) layers were incorporated for each branch. The distribution was preserved for the high accuracy of the regression. Experiments were conducted on the dataset with the 5 000 maize field images, covering a wide spectrum of the lighting conditions and growth stages. A sesame field dataset was also supplemented to verify the generalization. Ablation tests demonstrated that the collaborative integration of the modules yielded significant efficiency gains; Specifically, the LSPKI-YOLO model reduced the total parameter count by 27.2% and the floating-point operations by 20.7%, compared with the baseline YOLOv8s-Pose. While the inference speed simultaneously increased by 0.6 frames per second. In terms of the detection accuracy on the maize dataset, the model achieved a Mean Average Precision at 0.5 Intersection over Union (mAP₅₀) of 89.9% for the weed detection and 94.7% for the weed keypoint localization. In maize seedlings, the detection mAP₅₀ reached 94.6% with a keypoint localization mAP₅₀ of 96.7%. A comparison was also made on a broad range of the state-of-the-art models—including YOLOv7-Pose, YOLOv9-Pose, YOLOv10-Pose, YOLO11-Pose, YOLOv12-Pose, YOLOv13-Pose, Hyper-YOLO, and Mamba-YOLO. The highest accuracy metrics were achieved over the benchmark. Statistical t-tests (P<0.05) further confirmed there was the significance of the performance. Gradient-weighted Class Activation Mapping (Grad-CAM) visualization was performed to validate the interpretability. The network effectively focused on the global structural features of the large targets. While the precise attention was maintained on the local key regions of the small targets, thereby mitigating the false positives in the rest models. Moreover, the LSPKI-YOLO outperformed the YOLOv7 with the YOLOv10 variants in the cross-scenario validation within the sesame fields, indicating generalization over different crop types. Finally, the practical applicability was verified to deploy into a laser weeding robot on the Robot Operating System (ROS) that accelerated by TensorRT. Controlled indoor trials demonstrated a weed detection rate of 94.7% and a laser hit rate of 84.0%. In dynamic outdoor field trials under actual agricultural conditions, the robust adaptability was achieved in a recognition rate of 91.1% and a laser hit rate of 81.3%, while the crop damage rate was limited to 2.0%. The LSPKI-YOLO model successfully resolved the technological bottleneck between multi-scale feature enhancement and lightweight deployment in the specific context of the laser weeding. Both detection and localization accuracy were significantly improved to simultaneously reduce the computational load on the embedded hardware. The finding can provide a reliable and highly efficient visual perception. The field validation confirmed its practical utility for the high-precision, zero-herbicide weed control in modern smart agriculture. A viable technical foundation can offer for the broader application of the intelligent weeding robots in diverse crop environments.