optimizing one-stage crowd detection: an enhanced ssd with residual and trainable downsampling blocks

This paper presents an enhanced single shot multibox detector (ssd) architecture designed for highaccuracy small-object detection in dense crowd scenarios. traditional single-stage detectors often struggle with detecting small targets due to loss of fine-grained spatial information during downsampling. to address this, we propose an improved ssd (issd) that integrates residual feature blocks and trainable downsampling layers to preserve crucial low-level details. the proposed architecture employs multi-scale feature maps and adaptive default box configurations to enhance detection sensitivity. experimental results on voc and coco benchmarks demonstrate that issd achieves 78.68% map while maintaining real-time throughput of 89 fps, outperforming standard ssd300, ssd512, and faster r-cnn baselines. the system offers an optimal balance between speed and accuracy, suitable for real-time crowd analytics and surveillance applications.