LAROD-HD: Low-Cost Adaptive Real-Time Object Detection for High-Resolution Video Surveillance

Yanglin Pu, Chen Gao, Bo Liu, Si Liu, Shan Yang, Junhua Xiao, Shiliang Pu, Zhuhong You, Yanbin Wang

Abstract: An ideal video surveillance system should observe as many objects as possible while observing as many details of each object as possible simultaneously. This is hard to achieve since the resolution of most commercial off-the-shelf cameras is lower than 2k. High-resolution cameras, such as 8k, can observe enough details of each object while observing a wide FOV (field of view). However, high-resolution images bring high requirements for computing ability. We propose LAROD-HD, a low-cost adaptive object detection method, especially for small objects, in high-resolution surveillances with fixed-view. We utilize Bayesian rule to model the distribution of objects in the scene, so the high-value area can be processed with limited computing resources. In a new scene, LAROD-HD mines the regular pattern of object distribution to adapt to it automatically. Moreover, the adaptability is decoupled from the image, so the computational resource requirement is low and the domain shift problem is prevented. The experimental results on PANDA-8k dataset and UE4-8k dataset show that the proposed method achieved the best practical performance.