Automatic adaptation of a generic pedestrian detector to a specific traffic scene

Abstract

In recent years significant progress has been made learning generic pedestrian detectors from manually labeled large scale training sets. However, when a generic pedestrian detector is applied to a specific scene where the testing data does not match with the training data because of variations of viewpoints, resolutions, illuminations and backgrounds, its accuracy may decrease greatly. In this paper, we propose a new framework of adapting a pre-trained generic pedestrian detector to a specific traffic scene by automatically selecting both confident positive and negative examples from the target scene to re-train the detector iteratively. An important feature of the proposed framework is to utilize unsupervisedly learned models of vehicle and pedestrian paths, together with multiple other cues such as locations, sizes, appearance and motions to select new training samples. The information of scene structures increases the reliability of selected samples and is complementary to the appearance-based detector. However, it was not well explored in previous studies. In order to further improve the reliability of selected samples, outliers are removed through multiple hierarchical clustering steps. The effectiveness of different cues and clustering steps is evaluated through experiments. The proposed approach significantly improves the accuracy of the generic pedestrian detector and also outperforms the scene specific detector retrained using background subtraction. Its results are comparable with the detector trained using a large number of manually labeled frames from the target scene.

Adapting pedestrian detectors to new domains: A comprehensive review

“…As can be observed in Section 4.3, the overwhelming majority of the state-of-the-art research for domain adaptation of object detectors in videos use self-training in one form or another [29,30,31,32,33,36,37,38,40,42,11,10]. In order to adapt a generic pedestrian detector to a specific scene, a typical system would run the generic detector on some frames in a video, then score each detection using some heuristics and afterwards, add the most confident positive and negative detections to the original dataset for retraining.…”

“…In each iteration, positive and negative examples are collected by filtering with a variety of cues, added to the current dataset and a new classifier is trained. Figure taken from [11]. many situations.…”

“…As with other iterative self-training algorithms, the algorithm requires setting the threshold for selecting confident detections. Moreover, the system requires the use of multiple source domains which may not be feasible in Figure 12: An iterative self-training technique of Wang and Wang [11]. In each iteration, positive and negative examples are collected by filtering with a variety of cues, added to the current dataset and a new classifier is trained.…”

“…The method proposed by Wang and Wang [11] iteratively improves a generic pedestrian detector by selecting new confident examples to add to the current dataset for retraining at every iteration. In order to collect examples for each self-training iteration, their oracle is a combination of vehicle and pedestrian paths, multiple different cues such as bounding box locations and sizes, background subtraction, thresholds, filters and hier-archical clustering.…”

“…(a) CUHK Square dataset [10] (b) MIT Traffic dataset [11] (c) PETS 2009 dataset [12] Figure 2: Random samples from scene-specific pedestrian datasets (only pedestrians, i.e. positive examples, are shown).…”