| Pedestrian detection is a basic computer vision task,which has a wide range of applications in areas such as autonomous driving,video surveillance,and intelligent robots.Pedestrian detection aims at finding the location of pedestrians with the corresponding confidence in an image or video.Recent years have witnessed remarkable development in pedestrian detection,but in some complex scenes(small scale,occlusion,etc.),the performance is not satisfactory.To this end,this thesis carries out corresponding research works to solve the existing challenges.The specific research contents are as follows:(1)A multi-scale pedestrian detection algorithm based on a dynamic high-resolution network.The high-resolution network is good at the detection of small-scale pedestrians,but it usually fails to handle the scenario with a large scale variance.To solve this problem,this thesis proposes a pedestrian detector based on a dynamic high-resolution network.First,a soft conditional gate module is added to generate weights for weighted fusion between features with different resolutions.Then multiple parallel branches are constructed to generate scale-specific feature maps.The parameters of different branches are shared to avoid overfitting,except those in the soft conditional gate module.The experimental results show that,compared with the original high-resolution network,the dynamic high-resolution network proposed in this thesis has stronger generalization ability across scales,and thus significantly improves the detection accuracy of pedestrians at different scales.(2)An occluded pedestrian detection algorithm based on proposal relation-aware attention.Detecting occluded pedestrians is one of the most difficult tasks in pedestrian detection.To solve this problem,this thesis proposes to utilize region proposal features to generate the relation-aware attention and improve the feature representation.Specifically,this thesis mainly proposes two modules: one module uses the relationship between the feature nodes within each region proposal feature to compute the relation-aware attention map,thus generating more accurate region proposal features;another module measures the relationship between region proposal features,and supplements the information of nearby areas for each region proposal feature.The experimental results show that the method proposed can help the model extract features that are more robust to occluded pedestrians,thereby obtaining better performance.(3)A pedestrian detection algorithm in complex scenes based on long-tailed domain distribution.The above two works in this thesis and the existing works usually focus on one of the challenges and propose specific methods.However,different challenges may occur at a time simultaneously and change over time,making a specific method of limited usage in practice.Therefore,this thesis proposes a method that can detect pedestrians in various scenarios.First,the instance domain compactness is proposed to measure the distances among instances in the feature space,and handle hard cases from a novel long-tailed domain perspective.Specifically,a feature augmentation module is proposed to augment the tail instances in the feature space to balance the number of samples in the head and tail domains.In addition,a loss weighting module is constructed to utilize the instance domain compactness to generate loss weights.Thus,the loss value of each sample is re-weighted so as to allow the optimization to focus more on hard samples.The experimental results show that without any extra parameters at test time,the proposed method can achieve better performance across different challenging scenarios and is of high generalization ability. |
