Multi-Scale Feature Pyramid Network With Camera Artifact Augmentation for Pedestrian Detection

Nowadays, pedestrian detection is considered as an important area of applications in automated driving and surveillance systems. This paper presents a novel hybrid model termed as multi-scale feature pyramid network with camera artifact augmentation (MSFPN-CAA) that integrates Feature Pyramid Networks (FPN) with a fine-tuned YOLOv10 model for enhanced pedestrian detection. The model is augmented using various camera motion artifacts to simulate real-world dynamic environments. FPN is employed to extract multi-scale features. The fine-tuned YOLOv10 model predicts the bounding box regions. Non-maximum suppression (NMS) is applied to remove redundant detections. The paper presented the result with different batch size of the model, with different variants of YOLOV10. From the result analysis it was observed that batch size 16 is optimal batch size and variant YOLOV10n is the most accurate and YOLOV10s is the fastest version. Then result was compared with YOLOV9 and proposed fine- tuned YOLOV10 outperforms better as compared to YOLOV9. Then the result was presented for different camera artifacts. With all artifacts included, the model achieved a mAP of 72.9% with an average execution time of 4.5ms. Finally, the paper presented a comparative analysis with state-of-art models and achieved better performance.