Pedestrian Detection In Driver Assistance Systems

Pedestrian detection is an important part of the assisted driving system.Its main task is to determine whether pedestrians are included in the current road scene.If pedestrians are included,the specific coordinates of pedestrians need to be given,generally in the form of bounding boxes.Pedestrian detection in assisted driving systems needs to address the challenges posed by small target pedestrian detection.Small target pedestrians have lower resolutions,limited target features that can be extracted,and are more susceptible to missed detection due to noise interference.Pedestrian detection in the assisted driving system also needs to implement real-time pedestrian detection.At the same time,the accuracy of detection needs to be ensured on the basis of real-time performance.The current RefineDet algorithm has achieved good performance in object detection,but there are some deficiencies in dealing with small target pedestrian detection and real-time pedestrian detection.At the same time,as an algorithm based on the Anchor detection mechanism,there are still complex models and poor model migration capabilities.insufficient.This article has carried out related research on some existing problems of pedestrian detection,the main research work is as follows:1.Based on the RefineDet algorithm,by introducing a comprehensive feature enhancement module and a feature fusion module,it improves its feature extraction and fusion between different layers,thereby overcoming the problem of the RefineDet algorithm itself's insufficient ability to detect small target pedestrians.Experiments show that the improved algorithm has better detection performance in Caltech pedestrian dataset and BDD100 K dataset,especially for small target pedestrians in BDD100 K dataset.2.Because RefineDet has the advantages of both a one-stage detector and a twostage detector,so on the basis of the algorithm,by improving its basic network,adding Light-head structure blocks and feature fusion structures and introducing a series of efficient training The inference strategy constructs a lightweight and efficient real-time detection algorithm.Experiments show that the improved algorithm realizes the real-time pedestrian detection task on the BDD100 K data set,and at the same time guarantees a better detection accuracy.3.According to the implementation mechanism of Anchor,we can know that the commonly used Anchor-base series detection in order to improve the detection accuracy,the parameters are redundant and the model is more and more complicated,making it difficult to meet the requirements of efficient detection.This article will improve the basic network on the basis of the Anchor-free series of algorithms and introduce an adaptive sample selection strategy to solve the problem of selecting positive and negative samples in the detection algorithm,and finally achieve an efficient pedestrian detection algorithm.Experiments show that the improved algorithm has better detection results on the BDD100 K data set than the one-stage detection algorithm and most of the two-stage algorithms.At the same time,the detection results on the self-built data set also prove that its model migration ability is better than the Anchor-base algorithm.Can better adapt to different scenarios and data sets.