Small Object Detection Based on YOLOv5s in Advanced Driver Assistance System

With the rapid development of Advanced Driver Assistance Systems (ADAS), there has been a growing demand for accuracy of object detection in driving scenarios. The ambiguous features, along with the limited receptive field in the image, make it a challenging detection task for small objects. To address this issue, this paper proposed a novel small object detection approach based on an enhanced YOLOv5s model. The standard YOLOv5s algorithm struggles with detecting small objects accurately; hence, dynamic snake convolution is incorporated to enable the neural network to better extract features of the target. In addition, the SPPE module in YOLOv5s is replaced with the SPPELAN module to strengthen multi-scale feature extraction capabilities. Validated on a dataset derived from the public BDD 100K, the improved YOLOv5s algorithm achieves a 3% increase in detection accuracy compared to the original model. Notably, the detection accuracy for small objects, such as bicycles and motorcycles, improves significantly, by 3.3% and 8.6%, respectively. Experimental results confirm that the proposed method outperforms the standard YOLOv5s in the forward perception scenarios.