Research on obstacle detection algorithm based on object recognition

This paper delves into an obstacle detection algorithm that utilizes object recognition to enhance the precision of identifying small objects. By extending the YOLOv5s model with an additional layer dedicated to detecting small targets, we achieve multi-scale detection capabilities, thereby bolstering the recognition prowess for small obstacle targets. At the same time, the deformable convolution DCNv2 is combined to make full use of its ability to capture the features of small targets with large shape changes. To address the issue of redundant bounding boxes, we incorporate the Soft-NMS algorithm, which proficiently mitigates the inadvertent removal of bounding boxes, thereby enhancing both the accuracy and trustworthiness of the detection outcomes. Experimental findings underscore the remarkable performance improvement of the proposed algorithm, particularly in tasks involving the detection of small targets.