Unsupervised Hard Case Extraction Based on Image Perceptual Hash Encoding

The camera-based object detection algorithm is essential in autonomous driving. If the object is not detected when the vehicle is driving on the highway, it will cause a severe safety hazard. To evaluate and improve the object detection model, we propose a hard case extraction algorithm based on image perceptual hash encoding. We encode the object regions of each frame and then match them in adjacent frames. We optimize the search algorithm to achieve fast matching, improving efficiency by about ten times while ensuring accuracy. Then, we extract hard frames from a large number of unlabeled video frames, and the experimental results show that the accuracy is 92%. It is of great significance to evaluate and improve the object detection model and expand effective data.