<p>Object detection in complex street scenes faces significant challenges, including the prevalence of small-scale objects, severe occlusions, and imbalanced category distributions. Traditional YOLO-series models still exhibit limitations in detection accuracy and robustness under such conditions. To address these issues, this paper proposes an enhanced multi-scene adaptive detection model, BiFP-ATFL-YOLO11, built upon the YOLO11 framework. The model introduces two key improvements: first, a bidirectional feature pyramid network (BiFPN) with learnable fusion weights is integrated to strengthen multi-scale feature aggregation, enabling more flexible cross-scale feature interaction within the YOLO11 architecture. Second, a local adaptive mechanism is incorporated into the adaptive threshold focal loss (ATFL) to modulate gradient contributions according to prediction confidence under class-imbalanced training conditions. Comparative experiments conducted on the BDD100K real-world street scene dataset demonstrate that, under identical training and evaluation protocols, the proposed BiFP-ATFL-YOLO11 improves mAP@0.5 from 51.10% to 51.50% and mAP@0.5:0.95 from 30.4% to 30.6% compared with the YOLO11 baseline, while maintaining comparable precision and recall performance. These results indicate that the proposed architectural and loss-function enhancements provide a stable and reproducible performance gain without introducing additional training instability, supporting its applicability to real-time perception tasks such as autonomous driving and intelligent surveillance.</p>