Abstract:

Aiming at the problem that autonomous vehicles tend to confuse pedestrians and riders in the process of environmental perception， a joint detection method to effectively distinguish pedestrians and riders is proposed in this paper， with an improvement made based on faster region convolutional neural networks. Firstly， a subnetwork is added to extract the shape feature channel of images， which is then aggregated with the feature map generated by the backbone network， and the additional shape semantic channels are used for assisting the detector to distinguish the pedestrian and rider features. Then， a diversity region prediction unit is constructed to distinguish the distinct parts between pedestrians and riders and the suggested regions are divided into several parts with their confidence calculated， which is then used to determine the weight of part feature. Finally， the part features are aggregated with main?body feature and are sent to classifier for classification. The results of test on both pedestrian - rider data set built in this paper and open Caltech data set show that the improved detection method has relatively high detection accuracy， and the situations of the rider being mis?detected as pedestrian are effectively reduced.

Key words: Faster R?CNN, channel aggregation, diversity region prediction, pedestrian detection, rider detection