Optimized MobileNet plus SSD: a real-time pedestrian detection on a low-end edge device

One of the most fundamental challenges in computer vision is pedestrian detection since it involves both the classification and localization of pedestrians at a location. To achieve real-time pedestrian detection without having any loss in detection accuracy, an Optimized MobileNet + SSD network is proposed. There are four important components in pedestrian detection: feature extraction, deformation, occlusion handling and classification. The existing methods design these components either independently or in a sequential format, and the interaction among these components has not been explored yet. The proposed network lets the components work in coordination in such a manner that their strengths are improved and the number of parameters is decreased compared to recent detection architectures. We propose a concatenation feature fusion module for adding contextual information in the Optimized MobileNet + SSD network to improve the detection accuracy of pedestrians. The proposed model achieved 80.4% average precision with a detection speed of 34.01 frames per second (fps) when tested on the Jetson Nano board, which is much faster compared to standard video speed (30 fps). Experimental results have shown that the proposed network has a better detection effect during low light conditions and for darker pictures. Therefore, the proposed network is well suited for low-end edge devices.