Two-stage multiple object detection using CNN and correlative filter for accuracy improvement

Abstract

Recent advancements in tracking methods, particularly correlative-filter-based methods, can be utilized to support object detection to achieve the desired accuracy and speed. This is especially beneficial to aid a weak object detection model that results from the transfer learning concept trained on limited datasets, thus significantly reducing accuracy. Therefore, the limitation that comes from small datasets and training facilities needs to be tackled by a novel approach that is not restricted by computation time. The results of the proposed method (two-stage detection) are documented here. The first stage employs a convolutional neural network-based multi-object detector algorithm, and the results with low confidence scores are fed into the second stage for confirmation. A correlative filter is applied in the second stage; thus, low-confidence score results from the object detector algorithm are checked if the appearance is similar to the target object’s model appearance on the previous frame. Accuracy measurements were performed for both stages with different metrics to suit each algorithm goal. Image processing using CNN-based object detection yields bounding boxes with corresponding confidence scores, which show the probability of the object being detected as belonging to the target object class. The correlation filter tracking stage results are quantified using the novel movement distance score that will be explained. While the correlative filter plays a role in reducing false negatives by increasing detectability, the movement distance score ensures that the added detectability should not introduce more incorrect object tracking results. By applying the combination in the form of this two-stage method, a 10% enhancement in detectability (compared to a standalone object detection method) was observed in the various testing scenarios.