An Efficient Multi-Task Network for Pedestrian Intrusion Detection

In the growth of intelligent surveillance and computer vision, pedestrian intrusion detection (PID) has been widely applied in security, automatic driving and many other fields. Vision-based PID aims at detecting whether pedestrians invade an area-of-interest (AoI). For dynamic PID task, we need to identify the invading individuals from a large amount of pedestrian objects in a varying AoI scene, which is very challenging. We take the first attempt to solve this issue and propose a network named PIDNet to address it. In the PIDNet, we design an efficient dual-branch network, which combines region segmentation and pedestrian object detection. The final intrusion judgement is made by analyzing the position relation between a segmented AoI and a detected pedestrian. With the lightweight network design, PIDNet can detect pedestrian objects and segment AoI simultaneously, and make intrusion judgment efficiently. In order to achieve better intrusion detection performance, we further propose a netowrk with a cross-association strategy (cross-PIDNet) to improve the feature extraction capability of dual-branch network. In this strategy, we build the connection between detection branch and segmentation branch, avoiding the redundancy of non-critical intrusion features and enhancing the expression of critical intrusion features. Meanwhile, we design a multi-scale intrusion judgment module, which makes the judgment of intrusion state more robust. Since neither public datasets nor a benchmark is available in this direction, we set up a benchmark dataset and corresponding evaluation metrics. Experimental results indicate that our cross-PIDNet can reach 74.7%
$PID\_{A}cc$
and 50.8%
$PID\_{A}P$
.