With cameras as the main sensor, our AI-based algorithms detect, from the camera images, vehicles, pedestrians, cyclists, road lane lines, traffic signs, driver faces, and driver gestures. These detected objects are continuously tracked, and the algorithms analyze their movements and “intentions”. Based on this analysis, actions such as alerting drivers or even stopping the machines may be initiated.

From devices on vehicles to the cloud-based platform and the Vispect-specific statistical and analytical software, we have built a virtuous circle to enhance our algorithms. The footage and images, transmitted via 4G wireless from in-service vehicles (with the permission of the fleet owner and the driver, and with faces anonymized by the algorithm to protect personal information in compliance with local laws), serve as raw data for advanced Machine Learning and precision analysis. Consequently, the accuracy rates of Vispect's detection algorithms, measured by both True Positives and False Positives, has been greatly improved to a considerable high level.

In addition to classification, Semantic Segmentation has been achieved by our algorithms. By integrating these two technologies, we can not only detect various objects but also effectively distinguish them from their backgrounds. This capability provides a clearer understanding of complex driving environments - such as separating drivable areas from other parts of the image - thereby improving decision-making accuracy for Self-Driving.

Additionally, the Segmentation technology has significantly enhanced our ability to perceive the surrounding environment. Through real-time analysis of image captured by vehicle cameras, Vispect’s algorithms are equipped to respond rapidly to dynamic road conditions, ensuring both safety and operational efficiency.

Vispect's Convolutional Neural Networks (CNN) algorithm development team is diligently and intelligently enhancing the accuracy of our detection and segmentation models. Recognizing the importance of this long-term endeavor, we have developed a proprietary system known as the “AI Factory” to manage AI training datasets, labeling, and the processes of training and inference. Additionally, we have created a software tool called “Label-V” to label training data in accordance with the specifications of our AI models.

Assisted by these tools and with the input of over 100,000 videos per week—equivalent to 25,000,000 frames—we continuously deliver increasingly accurate algorithms to meet our users' needs. Furthermore, Vispect has assembled a professional team that utilizes our specialized Statistical and Analytical system to verify the accuracy of new algorithms. In an ongoing project, the analysis of 41,334 alarm videos revealed that the False Positive detection rate for a Vis G4-S system equipped with four top-down cameras was less than 0.5%, implying a True Positive rate exceeding 99.5%. The following are the results of recent statistics on the new algorithm's detection rate, with all samples sourced from real-world applications. These results demonstrate that the overall average accuracy of the algorithm has achieved 99.68%, with a False Positive rate of 0.32%.

Our experienced Hardware Development team is responsible for designing a diverse array of devices and systems that serve as the robust foundation for our AI algorithms. Our in-house development encompasses planning, performance evaluation, schematic design, protection design, adherence to industrial standards, circuit design, PCB layout, engineering prototypes, and New Product Introduction, culminating in the Mass Production phase. Our designs are crafted to comply with EMC, EMI, and environmental standards as stipulated by ISO and EN, or tailored to meet the specific requirements of car maker OEM fittings.
The multi-core System on Chip (SoC) serves as the brain of the Vispect system, with some models boasting more than one SoC to accelerate AI computations. To date, we have utilized SoCs from manufacturers such as ADT, Qualcomm, MTK, RK, and others. Microcontroller Units (MCU) are employed as the control core of our systems. We have incorporated MCUs from ST, NXP, Atmel , Microchip, and other manufacturers by far.

A team of approximately 20 experienced engineers comprises the Vispect software team. Their in-house responsibilities include analyzing and planning, firmware development, algorithm development, software development, software testing, road testing, and verification testing. Most Vispect SoCs utilize Android and Linux as their operating systems. The software development process strictly adheres to the V-Model methodology. White-box software testing is conducted on key or safety-critical modules, while black-box testing is applied to other modules. For high-level safety systems, the team follows the Safety Integrity Level 2 (SIL2) standard.

LiDAR, serving as a secondary sensor, performs exceptionally well in Vispect systems.
We utilize both single-plane and multi-plane LiDAR technologies. Our algorithms process the LiDAR point cloud data, which is then used for object detection through clustering, or fused with camera pixel data for advanced processing.
Since cameras have more pixels than LiDAR point clouds, they excel at object classification. However, due to its inherent characteristics, LiDAR offers significant advantages for various measurements, such as distance measurement, height detection, and deflection measurement.

Vispect systems are equipped with interfaces that facilitate the exchange of information with, interaction with, and control over vehicles and machinery. These interfaces enable the deployment of a variety of new functions tailored to specific needs.
For instance, the system can automatically stop a heavy Articulated Dump Truck when it detects a pedestrian too close to the vehicle. Similarly, it can initiate braking for a forklift when it senses a worker approaching from the direction it is moving towards.

To facilitate information exchange with other systems, Vispect systems generally offer a variety of interfaces, including CAN, RS-485, RS-232, Ethernet, T1, USB, video output, GPIO for transmitting and receiving ON/OFF signals, and 4G wireless connectivity to cloud-based data centers.
Interfaces are typically developed to meet OEM requirements. Most vehicles and machines utilize CAN interfaces. However, in some instances, to enhance safety, combined interfaces are employed. For example, an Automatic Train Protection system uses two Ethernet interfaces to exchange data with a Vispect system, serving as a mutual backup solution. Another example is an OEM forklift that uses one high/low voltage wire and a CAN interface to interact with a Vispect unit.