This study presents a Spatial Heatmap-Based Relevance Model that fuses the outputs of deep learning-based object detectors (such as YOLOv5 and Faster R-CNN). To address the limitations of single-model detectors, such as spatial ambiguity, occlusion sensitivity, and inconsistent confidence calibration, the model integrates spatial relevance modelling, confidence-adjusted fusion, and structured matching logic. A relevance heatmap is constructed from the training dataset to estimate class-specific spatial likelihoods across the image plane. During inference, each detected bounding box is assigned a relevance score based on its spatial alignment with the pre-computed heatmap, capturing the expected spatial distribution of object classes. These relevance scores are then used to rescale the original confidence values produced by each detector, allowing a more context-aware interpretation of the detection certainty. Subsequently, a Top N confidence product rule is applied to select the final detection outputs from both models. Evaluation in a road object detection dataset demonstrates consistent improvements in precision and F1 score compared to individual detectors and baseline fusion strategies. The proposed model remains modular and model-agnostic, making it compatible with various detection architectures that support spatial priors. The results highlight the benefits of incorporating spatial relevance learned into detection pipelines to enhance robustness in real-world scenarios.