YOLO-CS: A Lightweight Pedestrian Detection Model for Small Objects

In order for the self-driving system to ensure traffic safety, pedestrian detection is particularly important. But in the actual urban road environment, the situation is often more complicated, such as the distance of pedestrians from the car, and there are various obstacles on the road, which makes pedestrians look big and small, sometimes blocked. What is more troublesome is that there are many interference factors always make existing detection algorithms fail to maintain accurate recognition of multi-scale pedestrians in the real-time system. So, we propose an efficient and lightweight detection method based on the YOLOv8 model. Specifically, we incorporate the Convolutional Block Attention Module(CBAM) into the backbone part of the model, which effectively enhances the attention to salient regions of small objects during feature extraction while mitigating background noise interference. In the neck part of the model, we construct a SAC2f module based on Switchable Atrous Convolution(SAC), which is placed after the fusion of multi-scale feature maps. This design improves the model's overall multi-scale context modeling capability, enhancing multi-scale object detection. CBAM handles feature extraction while SAC2f improves fusion, boosting detection performance. Experimental results demonstrate our method achieves high detection accuracy and robustness while maintaining fast inference speed and low parameter count, meeting real-time requirements.