Method and system for detecting target difficult to distinguish

The invention discloses a method and system for detecting a target difficult to distinguish, and relates to the technical field of computer vision and Internet of Things, and the method comprises the following steps: S1, making a training data set, a verification data set and a test data set of a neural network; s2, calculating GIoU loss in a deep neural network loss function; s3, training and evaluating the deep neural model by using a simple heuristic super parameter adjustment method, and generating a deep neural network target feature extractor; s4, the deep neural network target feature extractor in the step 3 is converted into an ONNX file, and butt joint with a monitoring device in an occasion is achieved; s5, processing the original video frame by using a target detector; and S6, outputting detection result video frame sets Rt and Rt-n based on judgment on whether an alarm is given or not. The method can effectively solve the problems that an existing recognition method is low in target recognition pr