Walson Image Algorithm - Develop direction

• Traditional Algorithm 

At present, it is mainly used in two scenarios: alarm intrusion and single target tracking; Alarm intrusion algorithm: background modeling combined with Hungarian matching and Kalman filter to realize multi-target movement detection, tracking, and trajectory generation; The target speed, size, and other information are counted and analyzed, and the noise is filtered according to the statistical results; Support the movement detection and tracking of small targets; The combination of background modeling and deep learning classification realizes the function of motion detection + target classification; Single target tracking algorithm: support manual tracking and automatic tracking; Aiming at the problem that KCF is prone to target loss under the influence of occlusion, target scale change, and other factors, an outlier detection mechanism is added to improve the reliability of tracking.

• Deep Learning

The dual depth learning model recognition algorithm of visible light + thermal image is realized based on a convolutional neural network, which can detect, classify, track and judge the fusion of humans, vehicles, ships, UAVs, fireworks, and other targets, and support the multi-modal deployment of cloud side intercept framework / NVIDIA GPU framework server, side Atlas chip, end side Hisilicon chip and so on; Main technologies used: semi-automatic calibration based on background modeling / deep learning reasoning in repeated data scene, data enhancement such as tiling and mosaic for small targets, dynamic background modeling + depth classification algorithm framework fusion for thermal image scene, network tailoring optimization for specific computing platform, general algorithm standardization service for cross computing platform, etc.