Open AccessA GPU‐free real‐time object detection method for apron surveillance video based on quantized MobileNet‐SSD
Author(s) -
Lyu Zonglei,
Zhang Dan,
Luo Jia
Publication year - 2022
Publication title -
iet image processing
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.401
H-Index - 45
eISSN - 1751-9667
pISSN - 1751-9659
DOI - 10.1049/ipr2.12483
Subject(s) - computer science , object detection , quantization (signal processing) , overhead (engineering) , speedup , artificial intelligence , real time computing , computer vision , parallel computing , pattern recognition (psychology) , operating system
Object detection from apron surveillance video is facing enormous storage pressure and computing overhead. Large cloud server cluster is generally used and high‐speed network bandwidth is required, also equipped with powerful GPUs for computing support. The design of hardware‐friendly and efficient object detection model is challenging. This paper presents a compression method for outdoor apron surveillance videos, which is further combined with a lightweight detection model to make the inference process independent of GPU. First, the gray level variance of dynamic objects is leveraged to binarize the monitoring images, then an improved MobileNet‐SSD algorithm is proposed. Moreover, int8 quantization is performed and bit operations are designed to eliminate the floating‐point operation and it can simultaneously accelerate and compress CNN models with only minor performance degradation. Experiment results on a large‐scale dataset containing 22k monitoring images demonstrate that the compression ratio of quantized image can achieve up to 21 times, combined with quantized model, the detection on apron surveillance images can reach nearly 25FPS in a pure CPU environment, the mAP is 86.83%, and the model size is compressed to 600 kb. Significantly reduced computational complexity can be applied to embedded devices.