针对高速公路场景下抛洒物检测中因尺寸的类内差异大以及样本优化难度不均衡引发的漏检、误检问题，提出一种集成局部显著性-宽视野时空注意力与困难样本感知的对比损失函数优化框架。首先，利用孪生卷积神经网络提取参考图像与待检图像的外观特征。为提升多尺度抛洒物目标识别精度，设计了局部显著性-宽视野特征金字塔模块，该模块通过空间分块学习策略构建局部显著性特征金字塔，增强小尺寸目标的语义判别性；同时，利用多层次扩张卷积组合提取宽视野上下文信息，保障大尺寸目标的语义一致性。在此基础上，引入“抛洒突变引发时序变化”的先验知识，构建基于时序变化先验的时空注意力机制以聚焦抛洒物区域。最后，针对优化阶段的样本不均衡问题，提出困难样本感知对比损失函数，通过动态权重调整实现对易分样本的自适应抑制及对困难样本的深度挖掘。为了验证方法的有效性，本研究在真实高速公路场景下进行视频影像数据采集、标注，构建成标准数据集TDOD。实验结果表明：本文框架显著提升了小尺寸目标的局部显著性与大尺寸目标的语义连续性。相较于基线模型，该算法有效缓解了复杂场景下的精度瓶颈，核心评价指标调和均值F₁提高了4.7个百分点。可视化结果证实，该方法在抑制背景虚警、识别局部小目标及保持大尺寸目标轮廓完整性方面均表现出更强的鲁棒性。

To address the issues of missed detections and false alarms caused by significant intra-class scale variations and imbalanced sample optimization difficulty in highway dropped object detection, this paper proposes a detection framework integrating local saliency and wide-view feature spatiotemporal attention and a Hard Sample-Aware Contrastive Loss. First, a Siamese convolutional neural network is utilized to extract appearance features from both reference and test images. To enhance the recognition accuracy of multi-scale dropped objects, a Local-Saliency and Wide-View Feature Pyramid(LWFP) module is designed, which constructs a local saliency feature pyramid through a spatial partition learning strategy to improve the semantic discriminability of small-scale targets. Meanwhile, it simultaneously employs multi-level dilated convolution combinations to extract wide-view context to ensure the semantic consistency for large-scale targets. Building on this, a Spatio-Temporal Attention(STA) mechanism based on a temporal change prior is introduced to focus on dropped object regions by leveraging the prior knowledge that "dropped objects′ appearance inevitably triggers temporal changes". Finally, to mitigate the sample imbalance during the optimization phase, a Hard Sample-Aware Contrastive Loss function is proposed to adaptively suppress easy samples and intensify the mining of hard samples including missed debris and background false alarms through dynamic weight adjustments. Experimental results on the Traffic Dropped Object Detection(TDOD) dataset, constructed from real-world highway video data collection and annotation, demonstrate that the proposed framework significantly enhances the local specificity of small targets and the semantic continuity of large targets. Compared to the baseline model, the algorithm effectively alleviates accuracy bottlenecks in complex scenes, achieving a 4.7 percentage point improvement in the F₁-score. Visualization results confirm that the method exhibits superior robustness in suppressing background false positives, identifying localized small targets, and maintaining the contour integrity of large-scale objects.