Abstract
Temporal object detection has attracted significant attention, but mostpopular detection methods can not leverage the rich temporal information invideo or robotic vision. Although many different algorithms have been developedfor video detection task, real-time online approaches are frequently deficient.In this paper, based on attention mechanism and convolutional long short-termmemory (ConvLSTM), we propose a temporal single-shot detector (TSSD) forrobotic vision. Distinct from previous methods, we take aim at temporallyintegrating pyramidal feature hierarchy using ConvLSTM, and design a novelstructure including a high-level temporal unit as well as a low-level one(HL-TU) for multi-scale feature maps. Moreover, we develop a creative temporalanalysis unit, namely, attention-aware ConvLSTM (AC-LSTM), in which a temporalattention module is specially tailored for background suppression and scalesuppression while ConvLSTM temporally integrates attention-aware features. Anassociation loss is designed for temporal coherence. Finally, our method isevaluated on ImageNet VID dataset. Extensive comparisons on the detectioncapability confirm or validate the superiority of the proposed approach.Consequently, the developed TSSD is fairly faster and achieves an overallcompetitive performance in terms of mean average precision. As a temporal,real-time, and online detector, TSSD is applicable to robot's intelligentperception.