TESLA(TSLA.O)IN-DEPTH TRACKING REPORT：HOW TO VIEW THE PURE VISION ROUTE FOR TESLA AUTOPILOT

　　Core views:
　　Tesla, as a unique player among global vehicle manufacturers, has been insisting on the pure vision solution for autonomous driving, which has attracted a large number of followers because of its lower cost and faster commercialization path, despite many market concerns at the same time. Currently, the major market concern is that the pure vision solution has poor obstacle perception and is also susceptible to environmental factors, which can lead to safety issues. In our analysis, Tesla's algorithm design is based on “first principles” to realize the construction of 3D spatial bird’s eye view (BEV) and the integration of time series features with the help of BEV vision algorithms and raster networks, which greatly closes the gap with lidar solutions, while the environmental end of the problem can be avoided to the greatest extent through the driver’s subjective judgment. The systemic advantages brought by the integration of Tesla’s chips, algorithms and data, as well as the cost advantages of the pure vision solution, and a broader commercialization path built on stability and scalability, are likely to allow Tesla's autonomous driving technology to maintain a continuous lead in the global market and continue to cement its leading position in the global electric vehicle (EV) market. Affected by factors such as the Twitter incident and the market's concerns about short-term global car sales, Tesla's recent stock price performance has been weak. We continue to be optimistic about Tesla’s medium-/long-term investment value. With the stock trading at 41x/30x/21x 2022E/23E/24E non-GAAP PE, we advise investors to actively watch entry opportunities brought by the sharp pullback of Tesla’s share price in the short term.
　　Abstract:
　　Why the report: The market has concerns about Tesla's pure vision route, incl. obstacle recognition accuracy, robustness, and others.
　　Tesla has eliminated millimeter wave radars from all North American models since Feb 2022, switching to a pure vision solution for autonomous driving. Emerging EV makers NIO, Xpeng and Li Auto all adopt the "camera + radar + high precision map" scheme. Traditional carmakers like Volkswagen currently use the "camera + radar" scheme and are likely to carry lidar sensors in some models in the future. Because of the elimination of lidar sensors, the pure vision solution, by using cameras as the only source of data input, is more difficult to grasp the three-dimensional position information of obstacles. At the same time, the camera input is sensitive to light and other environmental effects and may cause difficulties in recognition in the case of backlighting and heavy snow cover. This report will systematically analyze and discuss the technical details, potential advantages and disadvantages of Tesla's pure vision route for autonomous driving, taking into account Tesla's latest technical developments, including relevant papers and technical details announced at the Tesla AI Day in the past two years.
　　Tesla's pure vision algorithm: Combining data engine capabilities and neural network models to help the algorithm iterate and bridge the gap with lidar solutions in perception.
　　1) Occupancy Network on the perception side: It joins the raster network to quickly classify the recognized objects according to the motion or static state and outputs them directly to the planning layer. Volumetric occupancy mapping is not a replacement for BEV, but a further extension of height direction based on BEV. After adding a dimension, volumetric occupancy mapping turns the 2D raster of BEV into 3D, which in turn generates occupancy features and replaces BEV features, and its performance has already been on par with or even surpassed lidar solutions in the high-speed moving environment. 2) Vector Lane for static recognition: solving complex urban reasoning in road route and road sign recognition. In the pre-processing, Tesla was the first to add the Map Component module, which uses the geometric topological relationship of lane lines as well as lane width and the number of lane lines in low-precision maps and integrates the code into the Vector Lane module. 3) Tesla's decision algorithm is based on the vector space constructed by its perception algorithm, and the overall decision is done with the help of incremental tree search. 4) Data Engine: emphasizing data extensibility to avoid dependence on high-precision maps. Tesla’s data engine features a self-contained closed loop, with a large amount of data collected by the fleet of standard autopilot hardware. This closed loop of data is undoubtedly the model of data-driven system application today, which is being copied by other manufacturers but difficult to surpass, as the core factor is Tesla's huge vehicle ownership and systematic engineering thinking.
　　The pros & cons of pure vision solutions: The gap in perception can be narrowed through algorithms & data iterations, while lidar solutions may get half the result with twice the effort because of synchronization & data-fetching problems.
　　1) The gap between pure vision and lidar solutions in perception can be significantly reduced through system engineering and technology. Tesla adopts holistic design and combines it with full-stack self-developed hardware and supporting software to optimize performance details to the extreme. 2) In multi-sensor fusion solutions, the camera is also an important part of the data input, and the camera in multi-sensor fusion solutions will also be seriously affected when the light source is affected, so it is extra work to do the fusion among multiple sensors. In addition, lidar detection enabled by emitting a beam is susceptible to the environment in extreme weather conditions, as it cannot function normally when the beam is blocked, and the multi-sensor design needs to rely more on the detection of millimeter wave radars in this case. But the low accuracy of millimeter wave radars itself makes it difficult for millimeter wave radars to serve as a good solution to this problem. In the face of extreme weather and other external conditions, decision-making still relies more on the driver's own judgment, regardless of the solution adopted.
　　Commercialization: Where pure visual solutions will shine the most.
　　1) The biggest advantage of the pure vision solution is the overall cost, as the hardware cost of Tesla's eight cameras is only about US$200, while the cost of a set of lidar sensors comes in the range of US$3,000-10,000. Troy Teslike data show that the current global penetration rate of Tesla’s Full Self-Driving software (Tesla FSD) was about 7.3% in 1Q22, compared to the peak of 35.7% in 4Q19, down by about 28.4ppts. The main reason is that the share of lower-priced models such as Model 3/Model Y in sales continues to increase. Based on Tesla's historical sales data, FSD penetration, and FSD prices, we estimate Tesla FSD's operating revenue reached US$830mn/870mn/940mn in 2019-2021, accounting for 3.4%, 2.8%, and 1.7%, respectively. 2) The second advantage of the pure vision solution is the uniformity and scalability of the system, which does not need to consider the synchronization problem of different perception units in a multi-sensory solution and the problem of which perception units should prevail when there is a difference in perception. This advantage will help Tesla's FSD to quickly expand its market space from small models to medium-sized cars, trucks and even humanoid robots. Looking ahead, we believe that the launch of Tesla Semi and Tesla Bot will, to a large extent, accelerate the commercialization of its self-driving software.
　　Potential risks:
　　Further deterioration of the pandemic worldwide; intensification of international trade conflicts; valuation fluctuations due to serious safety accidents of self-driving cars; disappointing progress in self-driving policy implementation; lower-than-expected development of the artificial intelligence technology, etc.; Tesla’s Berlin & Austin factories not progressing as expected, etc.; insufficient capacity of Tesla's power battery suppliers; insufficient grid load, etc.
　　Investment strategy:
　　We continue to be optimistic about the leading advantages of Tesla as the global EV leader in battery, battery management systems (BMS), autopilot algorithms, and intelligent driving data accumulation. The ideal scalability and economics in commercialization brought by the pure vision technology route are likely to support Tesla in achieving continuous leadership in autonomous driving and consistently strengthen its leading position in the global EV market. Despite Tesla's weak stock price performance recently due to the Twitter incident and market concerns about global auto sales, etc., we continue to be bullish on Tesla's medium-/long-term investment value. We maintain our 2022E-2024E revenue forecasts of US$82.7bn/118.3bn/154bn and non-GAAP net profit forecasts of US$14.0bn/19.1bn/27.2bn, implying 2022E/23E/24E non-GAAP PE of 41x/30x/21x at the current price. We advise investors to actively keep tabs on potential entry opportunities brought by the sharp pullback of Tesla’s share price in the short term.
【免责声明】本文仅代表第三方观点，不代表和讯网立场。投资者据此操作，风险请自担。

   【免责声明】本文仅代表第三方观点，不代表和讯网立场。投资者据此操作，风险请自担。