AI is everywhere. Artificial Intelligence AI–>Machine Learning ML–>Deep Learning DL–>Computer Vision CV, and the realization of AI intelligence functions in security mainly rely on computer vision to realize the understanding and recognition of images, speech and text. The AI function is mainly affected by the three factors of algorithm, computing power, and big data.AI function
AI functions in security include smart functions based on front-end cameras, smart functions based on back-end central platforms, or cloud-side integration.
Among them, based on the AI intelligent function of the front-end camera, there is a new concept, software-defined camera, that is, whether the intelligent function of the camera is mainly determined by software (SDC) or hardware, which has caused heated discussions and debates in the industry. My point of view: The market demand is determined, not by technology.
Common AI functions in the security field are divided into the following 4 categories:
Human body analysis: face recognition, posture recognition, human body feature extraction, etc.;
Vehicle analysis: license plate recognition, vehicle recognition, vehicle feature extraction, etc.;
Behavior analysis: target tracking and monitoring, abnormal behavior analysis, etc.;
Image analysis: video quality diagnosis, video summary analysis, etc.
If it is further subdivided, there are currently these smart functions:
Humanoid, animal detection
Snapshots, attributes, glasses, hairstyle, gender, facial expressions, etc.
People and traffic statistics
Vehicle, pedestrian road violation
Multi-dimensional perception (human expression, car, behavior analysis)
General smart function
Behavior analysis: cross-border detection, area intrusion detection, entry/exit area detection, wandering detection, people gathering detection, fast motion detection, parking detection, item left/take detection
Anomaly detection: scene change detection, audio anomaly (audio steep rise / steep fall detection, audio presence or absence detection), virtual focus detection
Statistics: over-line counting function
Support license plate capture, color, model, main brand, sub-brand, year model, etc. identification, machine non-human recognition, detection of forward or reverse driving vehicles, pedestrians and non-motor vehicles, automatic identification of vehicle license plates, you can capture no license plate Pictures of vehicles.
The computing power is the computing power of the processor. The general unit is TOPs, which is the abbreviation of Tera Operations Per Second. 1TOPs means that the processor can perform one trillion operations (10^12) per second.
Corresponding to GOPs (Giga Operations Per Second), MOPs (Million Operation Per Second). 1GOPs means that the processor can perform one billion times per second (10^9)
Operation, 1MOPs means that the processor can perform one million operations (10^6) per second.
In some cases, the power consumption of the processor is also an important factor to consider. Therefore, TOPs/W is also used as a performance index to evaluate the computing power of the processor. TOPs/W is used to measure how many trillion operations the processor can perform under the condition of 1W power consumption.
The computing power and power consumption of some common processors can refer to this statistic: Neural Network Accelerator Comparison
Lightspeeur2803 has an energy efficiency ratio of 24Tops/W, which is currently the AI chip with the highest energy efficiency ratio. Google’s TPU3.0 computing speed has reached 420TOPs.
In the security chip, Hi3516CV500 reached 0.5TOPs, Hi3516DV300/Hi3516AV300 reached 1.0TOPs, Hi3519AV100 reached 2.0Tops, and Hi3559AV100 reached 4.0TOPs (Huawei used this chip to make a 20TOPs police bayonet camera).
Rockchip’s RV1109 supports 1.2Tops, and RV1126 supports 2.0Tops.
The computing power is mainly determined by the performance of the AI chip. AI chips are mainly divided into four types: traditional CPU, GPU, semi-customized chips, and fully customized chips.
AI chip classification
Name Main Manufacturer Features
CPU Intel, AMD cannot process calculations in parallel
GPU NVIDIA, AMD have powerful parallel computing capabilities and floating-point computing capabilities. The GPU platform is very efficient in algorithm training. However, when processing a single input in inference, the advantages of parallel computing cannot be fully utilized.
FPGA Xilinx, Altera programmable array, semi-customized
ASIC TPU Google ASIC Tensor Processing Unit, Tensor Processor
NPU Cambrian, Apple NPU (Neural Network Processing Unit), Neural Network Processor
VPU Intel Vector Processing Unit vector processor
BPU Horizon Brain Processing Unit, brain processor
DPU Wave Computing Deep Learning Processing Unit, the deep learning processor
HPU Microsoft Holographics Processing Unit Holographic Image Processor
IPU Graphcore Intelligence Processing Unit
Algorithm (algorithm) is a series of well-defined specific calculation steps in mathematics and computer science. It is often used in calculations, data processing and automatic reasoning. As an effective method, an algorithm is used to calculate a function. It contains a series of clearly defined instructions and can be clearly expressed in a limited time and space.
Mathematicians and logicians in the early 20th century encountered difficulties in defining algorithms, until British mathematician Turing put forward the famous Turing thesis and proposed an abstract model of a hypothetical computer. This model is called a Turing machine. . The emergence of Turing machine solved the problem of algorithm definition, and Turing’s thought played an important role in the development of algorithm.
The core of the algorithm is to create an abstract model of the problem and a clear solution goal. After that, different modes and methods can be selected according to the specific problem to complete the design of the algorithm.
Common algorithms in machine learning include decision tree, random forest algorithm, logistic regression, SVM, naive Bayes, K nearest neighbor algorithm, K mean algorithm, Adaboost algorithm, neural network, Markov and so on.
Different algorithms may be used to solve different problems, or the same algorithm may be used. No certain algorithm is omnipotent, only the scope of application is different. It is important to choose the right algorithm for different scenarios.
There is no distinction between advanced and low-level algorithms. Quick and easy problem-solving is the foundation. Blindly pursuing complex algorithms is sometimes equivalent to “smashing mosquitoes with anti-aircraft guns.”
According to incomplete statistics, there are currently about 225 million surveillance cameras used in all walks of life across the country, and the amount of data generated in a medium-sized city in 90 days is about 36PB. How to use these isolated, disorderly, unstructured data, centralized management, and orderly and selective structured processing will be of extraordinary significance.
At present, a series of projects such as Xueliang Project, Safe City and Smart City have realized video surveillance networking in a certain area and centralized management of video data. These video big data can be used to help establish and train various algorithm models and improve algorithms. At the same time, the improved and optimized algorithm can be quickly added to the project practice and be tested in actual combat.
In the future, we can further cultivate and mine video big data to maximize its value.